Herod

Ma t h 6580
One must be sure that one has enabl ed sci ence to make a great advance i f one
i s to burden i t wi th many new terms and requi re that readers fol l ow research
that offers them so much that i s strange.
A. L. Cauchy
Li n ea r Algebr a , I n fi n i t e Di men si on a l Sp a ces,
a n d MAP LE
P REFACE
These notes evol ved i n the process of teachi ng a begi nni ng graduate
course i n Hi l bert Spaces. The fi rst edi ti on was si mpl y personal , handwri tten
notes prepared for l ecture. A copy was typed and gi ven to the students, for i t
seemed appropri ate that they shoul d have a statement of theorems,
exampl es, and assi gnments. Wi th each teachi ng of the course, the notes grew.
A text for the course has al ways been announced, but purchase of the
text has been opti onal . A text provi des addi ti onal readi ng, al ternate perspec-
ti ves, and a source of exerci ses. Whatever text was used, i t was chosen wi th
the i ntent of the course i n mi nd.
Thi s one quarter course was desi gned for sci ence and engi neeri ng
students. Often, as graduate students i n the sci ences and engi neeri ng
mature, they di scover that the l i terature they are readi ng makes references
to functi on spaces, to noti ons of convergence, and to approxi mati ons i n
unexpected norms.
The probl em they face i s how to l earn about these i deas wi thout
i nvesti ng years of work whi ch mi ght carry them far from thei r sci ence and
engi neeri ng studi es. Thi s course i s an attempt to provi de a way to understand
the i deas wi thout the students al ready havi ng the mathemati cal maturi ty
that a good undergraduate anal ysi s course coul d provi de. An advantage for
the i nstructor of thi s course i s that the students understand that they need to
know thi s subject.
The course does not devel op the i ntegrati on theory and noti on of a
measure that one shoul d properl y understand i n order to di scuss L
2
[0,1]. Yet,
these note suggest exampl es i n that space. Whi l e thi s causes some students
to feel uneasy wi th thei r unsophi sti cated background, most have enough
i ntui ti on about i ntegrati on that they understand the nature of the exampl es.
The success of the course i s i ndi cated from the fact that sci ence and
engi neeri ng students often choose to come back the next term for a course i n
real vari abl es and i n functi onal anal ysi s. The course has been provocati ve.
Even those that do not conti nue wi th more graduate mathemati cs seem to
feel i t serves them wel l and provi des an openi ng for future conversati ons i n
dynami cs, control , and anal ysi s.
I n thi s most current revi si on of the notes, syntax for MAPLE has been
added. Many sophomores l eave the cal cul us thi nki ng that the computer
al gebra systems are teachi ng tool s because of the system's abi l i ti es to graph,
to take deri vati ves, and to sol ve text-book di fferenti al equati ons. We hope,
before they graduate, students wi l l fi nd that these systems real l y are a "way
for doi ng mathemati cs." They provi de a tool for ari thmeti c, for sol vi ng
equati ons, for numeri cal si mul ati ons, for drawi ng graphs, and more. I t's al l i n
one program! These computer al gebra systems wi l l move up the curri cul um.
I n choosi ng MAPLE, I asked for a computer al gebra system whi ch i s
i nexpensi ve for the students, whi ch runs on a smal l pl atform, and whi ch has
an i ntui ti ve syntax. Al so, MATLAB wi l l run Mapl e syntax.
The syntax gi ven i n these notes i s not al ways the most effi ci ent one for
wri ti ng the code. I bel i eve that i t has the advantage of bei ng i ntui ti ve. One
hopes the student wi l l see the code and say, "I understand that. I can do i t,
too." Better yet, the student may say, "I can wri te better code!"
I t i s most i mportant to remember that these notes are about l i near
operators on Hi l bert Spaces. The notes, the syntax, and the presentati on
shoul d not i nterfere wi th that subject.
The i dea i n these notes that has served the students best i s the presen-
tati on of a paradi gm for a l i near operator on a Hi l bert Space. That paradi gm
i s ri ch enough to i ncl ude al l compact operators. For the student who i s i n the
process of studyi ng for vari ous exams, i t provi des a system for thi nki ng of a
l i near functi on that mi ght have a parti cul ar property. For the student who
conti nues i n the study of graduate mathemati cs, i t i s a proper pl ace to step
off i nto a study of the spectral representati ons for l i near operators.
To my col l eagues, Nei l Cal ki n and Eri c Bussi an, I say: The greatest
compl ement you can gi ve a wri ter i s to read what he has wri tten. To my
students: These notes are better because they read them and gentl y
suggested changes.
I am grateful to the students through the years who have provi ded
suggesti ons, correcti ons, exampl es, and who have demanded answers for the
assi gnments. I n ti me.... I n ti me....
James V. Herod
School of Mathemati cs
Georgi a Tech
Atl anta 30332-0160
Summer, 1997
Ta ble of Con t en t s
Sect i on Ti t le P a ge
1 A Decomposi ti on for Matri ces 1
2 Exp(tA) 5
3 Sel f Adjoi nt Transformati ons i n I nner-Product Spaces 7
4 The Gerschgori n Ci rcl e Theorem 11
5 Convergence 14
6 Orthogonal i ty and Cl osest Poi nt Projecti ons 18
7 Orthogonal , Nonexpansi ve, & Sel f-Adjoi nt Projecti ons 23
8 Orthonormal Vectors 27
9 The Fi ni te Di mensi onal Paradi gm 32
10 Bounded, Li near Maps from E to I C 35
11 Appl i cati ons to Di fferenti al and I ntegral Equati ons 40
12 The Si mpl e Paradi gm for Li near Mappi ngs from E to E 43
13 Adjoi nt Operators 45
14 Compact Sets 47
15 Compact Operators 50
16 The Space of Bounded Li near Transformati ons 53
17 The Ei genval ue Probl em 57
18 Normal Operators and The More General Paradi gm 61
19 Compact Operators and Orthonormal Fami l i es 64
20 The Most General Paradi gm:
A Characteri zati on of Compact Operators 67
21 The Fredhol m Al ternati ve Theorems 69
22 Cl osed Operators 72
23 Defi ci ency I ndex 74
24 An Appl i cati on: A Probl em i n Control 77
26 An Appl i cati on: Mi ni mi zati on wi th Constrai nts ??
25 A Reproduci ng Kernel Hi l bert Space ??
1
Linear Algebr a, Infinit e Dimensional Spaces, and MAPLE
Thi s cour se wi l l be chi efl y concer ned wi th l i near oper ator s on Hi l ber t
Spaces. We i ntend to pr esent a model , a par adi gm, for how a l i near
tr ansfor mati on on an i nner -pr oduct space mi ght be constr ucted. Thi s
par adi gm wi l l not model al l such l i near mappi ngs. To model them al l
woul d r equi r e an under standi ng of measur es and i ntegr ati on beyond what
a begi nni ng sci ence or engi neer i ng student mi ght know.
To set the patter n for thi s par adi gm, we fi r st r ecal l some l i near
al gebr a. We r ecal l , r evi ew, and r e-exami ne the fi ni te di mensi onal setti ng.
I n that setti ng, ther e i s the Jor dan Canoni cal For m. We pr esent thi s
decomposi ti on for matr i ces i n an al ter nate vi ew fr om the tr adi ti onal one.
The advantage to the r epr esentati on pr esented her e i s conceptual . I t sets a
patter n i n concepts, i nstead of i n "for m." We thi nk of pr ojecti ons and
ei genval ues. And, we must tur n to ni l potent matr i ces when pr ojecti ons and
ei genval ues ar e not enough.
Thi s i s how we begi n.
Sect ion 1: A Decomposit ion for Mat r ices
Definit ion A project ion i s a tr ansfor mati on P fr om E E such that P
2
= P.
Note that some texts al so requi re that P shoul d be non-expansi ve i n the
sense that | Px| | x| . An exampl e of a pr ojecti on that i s not non-
expansi ve i s P(x,y) = (x+y,0)
Definit ion A nilpot ent t ransform at ion i s a l i near tr ansfor mati on N fr om E
E for whi ch ther e i s an i nteger k such that N
k
= 0.
Examples
3 0-2
-1 1 1
3 0-2
i s a pr ojecti on fr om R
3
to R
3
and
-1 1 1
0 0 0
-1 1 1
i s a ni l potent
functi on fr om R
3
to R
3
.
Th eor em 1(Spectr al Resol uti on for A) I f A i s a l i near functi on fr om R
n
to
R
n
then ther e ar e sequences {
i
}
k
i =1
, {P
i
}
k
i =1
, and {N
i
}
k
i =1
such that each
i
i s
a number and
(a) P
i
i s a pr ojecti on, (b) N
i
i s ni l potent, (c)P
i
P
j
= 0 i f i j
(d) N
i
P
j
= 0 i f i j, (e) N
i
P
i
= N
i
(f) I =
1
k
P
i
,
and (g) A =
i
k
[
i
P
i
+ N
i
]
2
Outl i ne of Pr oof: We assume the Cayl ey-Hami l ton Theor em whi ch states
that i f A i n an nxn matr i x and D() i s the pol ynomi al i n defi ned by D() =
det(I -A), then the pol ynomi al i n A defi ned by maki ng the substi tuti on A =
sati sfi es D(A) = 0.
To constr uct the pr oof for the theor em, fi r st factor D() as
D() =
p=1
k
(-
p
)
m
p
,
wher e the p's ar e the zer o's of D wi th mul ti pl i ci ty m
p
. Now for m a par ti al
fr acti on decomposi ti on of D(): Let a
1
, a
2
, ... a
k
be functi ons such that
1
D()
=
p=1
k

a
p
()
(-
p
)
m
p
.
Two Exampl es
A
1
=
-1-2
3 4
and A
2
=
1 0 0
0 2 1
0 0 2
D
1
() = (-1)(-2) and D
2
() = (-1)(-2)
2
1
(-1)(-2)
=
-1
-1
+
1
-2

1
(-1)(-2)
2
=
1
-1
+
-+3
(-2)
2
I f q
p
() =
i p

(-
i
)
m
i
then
1 =
p=1
k
a
p
() q
p
() so that I =
p=1
k
a
p
(A) q
p
(A)
Two Exampl es Conti nued
1 = -1 (-2) + (-1) = (2-) + (-1)
1 = (-2)
2
+ (-+3)(-1)
I =
3 2
-3-2
+
-2-2
3 3
and I =
100
000
000
+
000
010
001
.
Cl ai m 1: Usi ng the Cal ey-Hami l ton Theor em i f P
j
= a
j
(A) q
j
(A) then P
i
P
j
=
0.
3
Cl ai m 2: P
j
i s a pr ojecti on si nce P
j
= P
j
I =
i =1
k
P
j
P
i
= P
j
2
.
Cl ai m 3: By the Cal ey-Hami l ton Theor em, i f
N
i
= a
i
(A)q
i
(A)(A-
i
I ) = P
i
(A-
i
I )
then N
i

m
i
= 0.
Cl ai m 4: N
i
P
i
= N
i
and N
i
P
j
= 0.
To see thi s note that N
i
P
j
= P
j
N
i
= P
j
P
i
(A-
i
I ) = 0
and N
i
P
i
= P
i
P
i
(A-
i
I ) = N
i
.
Fi nal l y, si nce I =
i

P
i
then
A =
i

P
i
A
=
i

P
i
(
i
I + A -
i
I )
=
i

[
i
P
i
+ N
i
].
Two Exampl es Fi ni shed
-1-2
3 4
= 1
3 2
-3-2
+2
-2-2
3 3
and

100
021
002
= 1
100
000
000
+ 2
000
010
001
+
000
001
000
.
Rema r ks
(1) The sequence {
i
}
k
i =1
i s the sequence of ei genval ues. I f x i s i n R
n
and
v
i
= P
i
N
i
m
i
- 1
(x),
then v
i
i s an ei genvect or for A i n the sense that
i
v
i
= A v
i
.
(2) For the ni l potent part
m
i
n.
4
I n fact, N
i
n
must be zer o for each i .
Assign men t Get the spectr al r esol uti on (or Jor dan Canoni cal For m) for the
matr i ces:
0 1
1 0
,
-3 2
1-2
,
0-1
1 2
, and
-5 1 3
1-2-1
-4 1 2
.
5
Sect ion 2: Exp(t A)
Often i n di ffer enti al equati ons -- both i n or di nar y and par ti al
di ffer enti al equati ons -- one can concei ve of the pr obl em as havi ng thi s for m
Z = AZ , Z(0) = c.
I f one can make sense of exp(tA), then sur el y thi s shoul d be the sol uti on to a
di ffer enti al equati on of the for m suggested. Fi ni te di mensi onal l i near
systems beg for such an under standi ng. Mor e i mpor tantl y, thi s
under standi ng gi ves di r ecti on for the anal ysi s of stabi l i ty questi ons i n
l i near , and nonl i near di ffer enti al equati ons.
Her e i s a r evi ew of the l i near , fi ni te di mensi onal case.
Theor em 2 I f P i s a pr ojecti on, t i s a number , and x i s i n {E, < , >}, Then
the sequence
S(n) =
i =1
n

t
i
P
i
i !
(x) =
i =1
n

t
i
i !
P(x)
conver ges i n E.
Recall that whatever nor m i s used for R
n
, i f A i s an nxn matr i x, then ther e
i s a number B such that i f x i s i n R
n
, then | Ax| B| x| . Mor eover , the
l east such B i s denoted | | A| | .
Definit ion exp(tA) =
i =0

t
i
A
i
i !
.
Cor ollar y 3 I f P i s a projecti on, exp(tP)(x) = (1-P)(x) + e
t
P(x).
Cor ollar y 4 Suppose that P and Q are projecti ons, PQ = QP = 0, and i f t i s a
number . Then
exp(tP+tQ) = e
t
P + e
t
Q + (1-P-Q).
Suggesti on of Pr oof. Wi th the supposi ti ons for P and Q, P+Q i s al so a
pr ojecti on. Thus, the pr evi ous r esul t appl i es.
Obser vation I f N i s ni l potent of order m then exp(tN) =
i =0
m-1

t
i
N
i
i !
.
Theor em 5 I f A i s a l i near tr ansfor mati on fr om R
n
R
n
and
A =
i
k
[
i
P
i
+ N
i
]
i s as i n Theor em 1, then
6
exp(tA) =
1
k
exp(
i
t)P
i
(
j=0
m
i
-1

t
j
N
i
j
j!
).
Suggesti on of Pr oof. Suppose that i s a number . Then
exp(tA) = exp(t) exp(t[A-I ]).
Suppose that A =
i
[
i
P
i
+N
i
].
Then P
i
exp(tA) = e
i
t
P
i
p=0

t
p
p!
(A-
i
I )
p
= e
i
t
P
i
p=0

t
p
p!
P
i
(A-
i
I )
p
si nce P
i
= P
i
2
.
= e
i
t
P
i
p=0

t
p
p!
N
i
p
.
Thus, exp(tA) = 1 exp(tA)
=
i
P
i
exp(tA)
=
i
e
i
t
P
i
p=0
m
i
-1

t
p
p!
N
i
p
.
Assign men t
Sol ve Z = , Z(0) = c, wher e A i s any of the matr i ces i n the pr evi ous
assi gnments.
7
Sect ion 3: Self Adjoint Tr ansfor mat ions in Inner -Pr oduct Spaces
I n these notes, we wor k i n an inner-product space. Often, we consi der the
space E as a col l ecti on of vector s over the compl ex (or r eal ) number s. We
take an inner product , or dot-product, to be defi ned on ExE and assume the
i nner pr oduct to have these pr oper ti es:
< x, y > = < y, x>*
< x+y, z > = < x, z > + < y, z>,
and < x, x> > 0 i f x 0.
Exa mples: Some i nner pr oduct spaces ever y young anal ysi sts shoul d know
are denoted R
n
, R
=
n
R
n
, L
2
, and L
2
[0,1]. We take R
n+1
R
n
i n a
cannoni cal way. Ever y i nner pr oduct space i s a nor med space for take the
norm to be defi ned by
| x| = <x,x> .
One shoul d note that not ever y nor med space ar i ses fr om an i nner pr oduct
space. I n fact, a char acter i zati on of those nor ms whi ch ar i se fr om an i nner
pr oduct i s that they sati sfy the par al l el ogr am i denti ty.
The di ffer ence i n a nor med space and an i nner pr oduct space i s an
i mpor tant noti on. I t woul d be wel l to have some exampl es of i nner pr oduct
spaces and nor med spaces to bui l d i ntuti on. We pr esent exampl es i n cl ass.
The best known i nequal i ty i n an i nner pr oduct space i s the Cauchy
Schwar tz i nequal i ty.
Theor em 6: I f x and y are i n E then | < x, y>| | x| | y| .
Hi nt for pr oof: Consi der h(t) = x + <x,y> y t and | h(t)|
2
.
Li near tr ansfor mati ons that ar e sel f adjoi nt have a si mpl e str uctur e.
We begi n wi th these.
Definit ion I n an i nner pr oduct space {E,< , >}, the l i near tr ansfor mati on A
on E i s self adjoint i f <Ax,y> = <x,Ay> for al l x and y i n the domai n of A.
Exa mples: Her e ar e thr ee exampl es to r emi nd you that the noti on of "sel f-
adjoi nt" i s per vasi ve.
(a) I f A i s
0 1
1 0
then A i s sel f-adjoi nt i n R
2
wi th the usual dot pr oduct.
(b) I f K(f)(x) i s
0
1
cos(x-y) f(y) dy then K i s sel f-adjoi nt on L
2
[0,1] wi th the
usual dot pr oduct.
8
(c) I f A(f) = (pf ) + qf, wi th p and q smooth, then A i s sel f-adjoi nt on the
col l ecti on of functi ons i n L
2
[0,1] havi ng two deri vati ves and that are zero at
0 and at 1. A useful i denti ty to under stand thi s exampl e i s
(pf )g f(pg ) = [ p(f g fg)].
Definit ion: Number s and vectors v are ei genval ues and ei genvect ors,
r especti vel y, for the l i near tr ansfor mati on A i f Av = v.
Theor em 7: I f A i s sel f-adjoi nt then al l ei genval ues ar e r eal and
ei genvector s cor r espondi ng to di ffer ent ei genval ues ar e or thogonal .
Suggesti on of Pr oof. Suppose that x0 and that Ax=x. We hope to show
that = *:
<x,x> = <Ax,x> = <x,Ax> = *<x,x>.
Fi nal l y (-)<x,y> = <Ax,y> - <x,Ay> = 0, so that <x,y> = 0.
Cor ollar y 8: I n the decomposi ti on of Theor em 1, i f A i s sel f adjoi nt then
each N
i
= 0.
Suggesti on of a pr oof. Let m be the smal l est i nteger such that N
m
= 0 and
suppose that m 1. Then, for each x i n E and use the fact that each N i s a
pol ynomi al i n A to see that N i s sel f adjoi nt i f A i s. Consequentl y,
<N
m-1
x, N
m-1
x> = <N
m-2
x, N
m
x> = 0.
Assign men t
(3.1) Gi ve a poi nt i n L
2
that i s not i n R
.
(3.2) Constr uct a pr oof of the Cauchy Schwar tz i nequal i ty i n a gener al i nner
pr oduct space.
(3.3) Show that

-.5 .5
.5-.5
i s sel f adjoi nt wi th the usual dot pr oduct. Fi nd i ts Jor dan For m. Repeat
thi s for
-.5 .5 0
.5-.5 0
0 0 0
.
(3.4) Suppose that A i s sel f-adjoi nt and has spectr al r esol uti on gi ven by
9
A =
i =1
k

i
P
i
.
Suppose al so that
1
= 0 and al l other 's ar e not zer o. Recal l i ng that i f e i s
i n E then
e =
i =1
k
P
i
(e),
show that P
1
(e) i s i n the nul l space of A and that

i =2
k
P
i
(e)
i s i n the r ange of A. I n fact, fi nd u such that
A(u) =
i =2
k
P
i
(e).
(3.5) Fi nd the ei genval ues and ei genvector s for the matr i ces i n the
Assi gnment of Secti on 1. Rel ate thi s to the Jor dan For m for sel f adjoi nt
matr i ces.
(3.6) Veri fy that

1
1
0
,
1
-1
0
, and
0
0
1
ar e or thogonal . Cr eate A such that these thr ee vector s ar e ei genvector s and
0,-1, and 2 ar e ei genval ues for A. (Hi nt:
Ax = 0 <x,v
1
> v
1
/| v
1
| 1 <x, v
2
> v
2
/| v
2
| + 2 <x, v
3
> v
3
/| v
3
| .)
(3.7) Let
K[f](x) =
0
1
cos((x-y)) f(y) dy
on L
2
[0,1]x[0,1]. Fi nd the ei genval ues and ei genfuncti ons for K. Show that
the ei genfuncti ons ar e or thogonal .
10
Sect ion 4: The Ger schgor in Cir cle Theor em
I f al l the ei genval ues of A ar e i n the l eft-hal f pl ane -- have negati ve
r eal par t -- then sol uti ons for
Z = AZ
col l apse to zer o. I f at l east one ei genval ue i s i n the r i ght-hal f pl ane, then
some sol uti on gr ows wi thout bound. Consequentl y, i t i s ver y i mpor tant to
know wher e the ei genval ues of A ar e, even i f the matr i x i s such a mess that
the ei genval ues ar e har d to fi ne. You can guess that peopl e have tur ned
thei r attenti on to l ocati ng ei genval ues. The next theor em i s a cl assi cal
r esul t. Mor e r ecent r esul ts ar e i ncl uded i n thi s secti on as a ci ted r efer ence.
Theor em 9 (Ger schgor i n's Ci r cl e Theor em) I f A i s a matr i x and
S=
m
{z: | A
mm
-z|

j m

| A
j m
| },
then ever y ei genval ue of A l i es i n S.
Suggesti on for Pr oof.
Lemma : I f A i s a matr i x and
| A
mm
| >
i m

| A
i m
|
for al l m, then A i s i nver ti bl e.
Here's how to see thi s: Suppose A i s not i nverti bl e. Thus 0 = det(A) =
det(A
T
). Let x be such that A
T
x = 0 and x 0. Let m be such that | x
m
| =
max
i
| x
i
| . Then
0 =(A
T
x)
m
=
i

A
T
mi
x
i .
Or , A
mm
x
m
=
i m

A
T
mi
x
i
=
i m

A
i m
x
i
.
Hence, | A
mm
| | x
m
| = |
i m

A
i m
x
i
| | x
m
|
i m

| A
i m
| .
Pr oof of Theor em 9: Suppose that i s an ei genval ue of A not i n S. Then
| -A
mm
| >
i m

| A
i m
| .
Thus, det(I -A) 0 and can not be an ei genval ue. Thi s i s a contr adi cti on.
11
Rema r k: I t i s easy to see that the Ger shgor i n Ci r cl e Theor em mi ght be
i mpor tant for deter mi ni ng the behavi or of systems of di ffer enti al equati ons.
For exampl e, as noted i n the begi nni ng of thi s secti on, i f al l ei genval ues of
the matr i x A ar e i n the l eft hal f of the compl ex pl ane, then sol uti ons of the
di ffer enti al equati on
Z = AZ
have asymptoti c l i mi t zer o.
As a r esul t of the i mpor tance of these i deas, consi der abl e wor k has
been done to i mpr ove the bounds. One i mpr ovement can be found i n the
fol l owi ng:
Davi d G. Fei ngol d and Ri char d S. Var ga, Block Diagonally Dominant
Mat rices and Generalizat ions of t he Gerschgorin Circle Theorem, Paci fi c
J. Math., 12 (1962) 1241-1249. (QA1.P3.)
I n that paper , the author s consi der bl ock matr i ces down the di agonal and
state the i nequal i ti es i n terms of a mat ri x n orm.
Definit ion: I f A i s a matr i x on R
n
wi th nor m | x| =
i
x
i
2
, then the 2-norm
of A i s defi ned by
| | A| |
2
= sup (
| Ax |
| x|
, | x| 1).
Ma t r ix n or ms:
I f the nor m for the space i s gi ven by
| x|
1
=
i =1
n
| x
i
| then | | A| |
1
= max
p

i =1
n
| A
i ,p
| -- the maxi mum col umn
s u m.
| x|
2
=
p=1
n
| x
i
|
2
then | | A| |
2
= the squar e r oot of the maxi mum
ei genval ue of the matr i x A (tr anpose A).
| x|

= max
i
| x
i
| then | | A| |

= max
p

i =1
n
| A
p,i
| -- the maxi mum r ow
s u m.
Example: Consi der the matr i x
12
A =
4 -2 -1 0
-2 4 0 -1
-1 0 4 -2
0 -1 -2 4
.
The r esul t pr esented i n these notes establ i sh that al l ei genval ues l i e i n the
di sk | - 4| 3. The i mpr oved r esul t i n the r efer ence ci ted above show that
al l ei genval ues l i e i n the di sks | - 6| 1 and | - 2| 1. Ei genval ues ar e
actual l y 1, 3, 5, and 7.
Assign men t
(4.1) Pr ove Theor em 9 for r ow sums i nstead of col umn sums.
(4.2) Let A =
2 0 1
0 1 0
1 0 2
.
Show that the ei genval ues of A sati sfy | -2| 1 or = 1. Gi ve the
ei genval ues for A. Get the Jor dan for m for A.
14
Sect ion 5: Conver gence
I n many appl i cati ons of mathemati cs, one can onl y make
appr oxi mati ons. I f you thought of maki ng successi ve appr oxi mati ons, you
woul d hope to be maki ng a sequence whi ch conver ges. The questi on i s, i n
what sense woul d the appr oxi mati on be good? I n what sense woul d the
sequence of appr oxi mati on get cl ose? Thi s secti on sets a fr amewor k for
under standi ng these questi ons by consi der i ng var i ous for ms of
conver gence.
Suppose that {E, < , >} i s an i nner pr oduct space of functi ons on an
i nter val . We wi l l di scuss four methods of conver gence: str ong, weak,
poi ntwi se, and uni for m.
Definit ion
A sequence of poi nts {x
p
} converges to y st rongly i f l i m
p
| x
p
- y| =0.
A sequence of poi nts {x
p
} converges to y weak ly i f, for each h i n E,
l i m
p
< x
p
- y, h> = 0.
A sequence of functi ons {f
p
} converges to g point wise on C i f for each x i n C,
l i m
p
f
p
(x) = g(x).
A sequence of functi ons {f
p
} converges to g uniformly on C i f
l i m
p
max| f
p
(x) - g(x)| = 0.
Examples
(1) Str ong conver gence i mpl i es weak conver gence, but not conver sel y.
Str ong conver gence Weak conver gence:
Suppose h i s i n E and {x
p
} converges to y st rongly . By the Cauchy-Schwar tz
i nequal i ty | < x
p
- y, h>| | x
p
- y| | h| .
To see that weak conver gence does not i mpl y str ong conver gence, we need
an exampl e. One exampl e comes fr om exper i ences wi th Four i er Ser i es.
Suppose h L
2
[0,1]. Ther e i s a sequence a
p
wi th l i m
p
a
p
= 0 so that
h(x) =
p=0
a
p
si n(px).
Thus, the cl ai m i s that
p
defi ned by
p
(x) = si n(px), conver ges weakl y to
zero but not strongl y. We have that for each h, = a
p
0 and
|
p
- 0|
2
=
0
1
si n
2
(px) dx =
0
1

1+cos(2px)
2
dx =
1
2
.
(2) Uni for m conver gence i mpl i es poi ntwi se conver gence, but not conver sel y.
15
That uni for m conver gence i mpl i es poi nt wi se conver gence i s not so har d.
To be assur ed that the i mpl i cati on does not go the other way, consi der
f
n
(x) = nxe
-n x
and note f
n
(1/n) = 1/e.
Fi gur e 5.1 i s an i l l ustr ati ve gr aph. I t shows thi s sequence conver gi ng
poi ntwi se, but not uni for ml y.
Fi gur e 5.1
I t's ti me to say what a Banach space i s and what a Hi l ber t space i s.
Definit ions: A sequence {x
p
} con verges i f
l i m
m,n
| x
m
- x
n
| 0.
A space i s com pl et e i f ever y sequence that conver ges has a l i mi t i n the
space. A Banach space i s a compl ete, nor med l i near space. A Hilbert space
i s a compl ete, i nner pr oduct space.
Another noti on that i s cr i ti cal i n the r emai nder of these notes and i s l i kel y
i n the vocabul ary of the reader al ready i s that of a closed set . A cl osed subset
of a Hi l bert space or Banach space i s a set wi th the property that i f {x
p
} i s a
sequence wi th val ues i n the set and l i m
p
x
p
= y, then y i s al so i n the set.
Ther e ar e al ter nate char acter i zati ons: cl osed sets ar e sets that contai n al l
thei r boundar y poi nts, ar e sets whose compl ements ar e open.
Assign men t :
(5.1) Show that i f | x
n
| | L| and x
n
L weakl y, then x
n
L str ongl y.
(Hi nt: Show that | x
n
-L|
2
|L|
2
-2| L|
2
+| L|
2
.)
(5.2) Show that uni for m conver gence on [0,1] i mpl i es str ong conver gence i n
L
2
[0,1], but not conversel y.
16
(5.3) Show that poi ntwi se conver gence on [0,1] does not i mpl y str ong
conver gence i n L
2
[0,1] and that str ong conver gence i n L
2
[0,1] does not i mpl y
poi ntwi se conver gence on [0,1].
(5.4) Di scuss the natur e of the conver gence of

(-1)
n+1

si n(nx)
n
on the i nterval [0,2]. (Hi nt: her e i s the gr aph of
1
4
(-1)
n+1

si n(nx)
n
.
x
6 5 4 3 2 1 0
1.5
1
0.5
0
-0.5
-1
-1.5
(5.5) Suppose that l i m
p
x
p
= y and l i m
p
u
p
= v.
(a) Show that l i m
p
< x
p
, u
p
> = < y, v>.
(b) Show that l i m
p
| x
p
| = | y| .
(c) Let {
p
} be an i nfi ni te maxi mal or thonor mal sequence, x
p
=
p
, and
u
p
=
p
/2. Show that the weak l i mi t of x
p
and u
p
i s zero, but
l i m
p
< x
p
, u
p
> 0.
MAP LE Rema r k: Gr aphs of the fi r st four functi ons i n a sequence that
conver ges poi ntwi se, but not uni for ml y ar e made wi th MAPLE i n a r ather
i ntui ti ve way.
> f 1: =x- >x* exp( - x) ; f 2: =x- >2* x* exp( - 2* x) ; f 3: =x- >3* x* exp( - 3* x) ;
f 4: =x- >4* x* exp( - 4* x) ;
> pl ot ( {f 1( x) , f 2( x) , f 3( x) , f 4( x) }, x=0. . 6) ;
You mi ght pr efer thi s syntax and gr aph.
> f : =( n, x) - >n* x* exp( - n* x) ;
> pl ot 3d( f ( n, x) , n=1. . 4, x=0. . 6, axes=NORMAL, or i ent at i on=[ - 10, 50] ) ;
17
The space C[-1,1] of conti nuous functi ons i n L
2
[-1,1] i s an exampl e of
an i nner pr oduct space that i s not compl ete. When asked to show i t i s not
compl ete, a novi ce mi ght dr aw a sequence of functi ons i n a manner
suggested wi th thi s syntax:
> pl ot ( {[ x, 1- 2* x, x=0. . 1/ 2] , [ x, 1- 3* x, x=0. . 1/ 3] , [ x, 1- 4* x, x=0. . 1/ 4] ,
[ x, 1+2* x, x=- 1/ 2. . 0] , [ x, 1+3* x, x=- 1/ 3. . 0] , [ x, 1+4* x, x=- 1/ 4. . 0] },
x=- 1. . 1) ;
Thi s i s not a convi nci ng pi ctur e for L
2
[-1,1] i s mor e compl i cated than
suggested by that pi ctur e. To see thi s, establ i sh that the sequence suggested
by that pi cture converges to zero by seei ng that
-1
1
[f
n
(x) - 0]
2
dx = 2
0
1/n
(1-nx)
2
dx = 2/(3n).
> i nt ( ( 1- n* x) ^2, x=0. . 1/ n) ;
Thi s i ntegr al goes to zer o as n i ncr eases.
How does one make an exampl e of a sequence i n C[-1,1] that
conver ges but has a l i mi t not i n C[-1,1]? Her e's a suggesti on.
> pl ot ( { [ x, Pi * ( 1- x) / 2, x=0. . 1] , [ x, - Pi * ( 1+x) / 2, x=- 1. . 0] ,
[ x, sum( si n( p* Pi * x) / p, p=1. . 5 ) , x=- 1. . 1] }) ;
That gr aph suggests a good i dea for an exampl e. One must consi der
whether
p

si n(px)/p
conver ges i n C([-1,1]). Thi s br i ngs up i ntegr ati ng the sum of the squar es of
the ter ms. Note the cal cul us her e
> si mpl i f y( i nt ( si n( p* Pi * x) * si n( q* Pi * x) , x = - 1. . 1) ) ;
Thi s i s zer o i f p and q ar e di ffer ent i nteger s. On the other hand
> subs( si n( p* Pi ) =0, i nt ( ( si n( p* Pi * x) / p) ^2 , x=- 1. . 1) ) ;
We are ready to see the seri es i s Cauchy. We see that the l i mi t of the seri es
i s the odd extensi on of (1 - x)/2.
> i nt ( ( Pi * ( 1- x) / 2) ^2, x=0. . 1)
- 2* sum( i nt ( Pi * ( 1- x) * si n( p* Pi * x) / ( 2* p) , x=0. . 1) , p=1. . n)
+ sum( i nt ( ( si n( p* Pi * x) / p ) ^2, x=0. . 1) , p=1. . n) ;
> l i mi t ( subs( si n( p* Pi ) =0, " ) , n=i nf i ni t y) ;
Thi s l i mi t i s zer o and we have that the sequence has l i mi t not i n C[-1,1].
18
Section 6: Or thogonality and Closest Point Pr ojections
We now wi l l study the noti ons of or thogonal i ty and or thogonal
pr ojecti ons. Thi s wi l l expand our noti on of what a pr ojecti on i s. Cr i ti cal i n
maki ng up "cl osest poi nt " pr ojecti ons i s that the r ange of the pr ojecti on
shoul d be convex. We exami ne thi s geometr i c noti on i n thi s secti on.
The i denti ty of part 1 of the next theorem i s cal l ed the Pyt hagorean
Ident it y and the i denti ty of part 2 i s cal l ed the Parallelogram Ident it y.
Theor em 10 (1) I f <x, y> = 0, then | x+y|
2
= | x|
2
+ | y|
2
.
(2) I f x and y ar e i n E, then | x+y|
2
+ | x-y|
2
= 2| x|
2
+ 2| y|
2
.
What fol l ows i s a revi ew of the noti on of convexi ty for use i n the
expl or ati on of pr ojecti ons.
Definit ion A set C i s convex i f for al l x and y i n C and al l number s t i n [0,1],
tx + (1-t)y i s i n C.
REMARK: The next r esul t gi ves a char acter i zati on of cl osest poi nt
pr ojecti on i n ter ms of the nor m and al so i n ter ms of the dot pr oduct.
Th eor em 11 I f C i s a cl osed, convex subset of the Hi l bert space {E, < , >}, x
0
E, and = i nf{| | x
0
- c| | , c C} then there i s one and onl y one y
0
i n C
such that | | x
0
- y
0
| | = . Mor eover , these ar e equi val ent:
(a) z i s i n C and Re<x
0
-z, c-z> 0 for al l c i n C, and
(b) z = y
0
.
Outl i ne of Pr oof: Suppose that C i s a cl osed convex set and that x
0
i s i n E.
Let = i nf{| x
0
-y| : yC}
Fi r st we show ther e i s a poi nt y
0
i n C such that | x
0
- y
0
| = . To do
thi s defi ne a sequence {u
p
} i n C by | u
p
- x
0
| < +
1
p
. Note that {u
p
}
p=1
conver ges for | u

n
- u
m
|
2
= | (u
n
-x
0
) - (u
m
- x
0
)|
2
= 2| u
n
-x
0
|
2
+ 2| u
m
-x
0
|
2
- 4 | (u
n
+u
m
)/2 - x
0
|
2
by the par al l el ogr am
i denti ty. Because C i s convex, (u
n
+u
m
)/2 i s i n C and thi s l ast l i ne does not
exceed 2| u
n
-x
0
|
2
+ 2| u
m
-x
0
|
2
- 4
2
whi ch goes to zer o as n and m goes to
i nfi ni ty. Hence, the sequence {u
p
} conver ges str ongl y and has l i mi t some
poi nt y
0
i n E si nce E i s compl ete. I t i s i n C si nce C i s cl osed.
Now we show ther e i s onl y one cl osest poi nt to x
o
i n C. Suppose that
| x
0
- z | = = | x
0
- y
0
| .
Then
| y
0
- z|
2
= | (y
0
- x
0
) - (z-x
0
)|
2
= 2| y
0
-x
0
|
2
+ 2| z-x
0
|
2
- 4 | (y
0
+z)/2 - x
0
|
2
= 0.
Thus, y
0
= z.
19
What fol l ows i s an under standi ng of the char acter i zati on of the
cl osest poi nt usi ng the i nner pr oduct.
These ar e equi val ent:
(a) z i s a poi nt i n C and Re<x
0
-z,c-z> 0 for al l c i n C, and
(b) z i s y
0
.
ab Suppose that z has thi s property (a). We hope to show that z = y
0
.
Because of the uni queness of thi s cl osest poi nt, we woul d be content to show
that | x
0
-z| = | x
0
-y
0
| .
We know that | x
0
-y
0
| | x
0
-z| .
0 | x
0
-z|
2
- | x
0
-y
0
|
2
= | x
0
-z|
2
- | (x
0
-z)+(z-y
0
)|
2
= | x
0
-z|
2
- [| x
0
-z|
2
- 2 Re<x
0
-z,y
0
-z> + | z-y
0
|
2
]
= 2 Re<x
0
-z,y
0
-z> - | z-y
0
|
2
0.
Ther efor e, | x
0
-z|
2
= | x
0
-y
0
|
2
.
Now we show that b a by showi ng that y
0
has thi s dot pr oduct
char acter i zati on. Recal l that the set C i s convex so that i f c i s i n C then so i s
tc+(1-t)y
0
i n C. The fi r st i nequal i ty hol ds because y
0
i s the cl osest poi nt i n C
to x
0
:
0 | x
0
-y
0
|
2
- | x
0
-[tc+(1-t)y
0
]|
2
= | x
0
-y
0
|
2
- | [x
0
-y
0
] + t(y
0
- c)|
2
= | x
0
-y
0
|
2
| x
0
-y
0
|
2
+2t Re<x
0
y
0
, c-y
0
> - t
2
| y
0
- c|
2
for al l 0 < t 1.
Ther efor e, Re<x
0
-y
0
, c-y
0
> 0.
REMARK. Thi s r esul t l eads to two i r r esi stabl e questi ons:
(1) I n a Banach Space that i s not a Hi l bert Space and gi ven a poi nt, i s there
a uni que cl osest poi nt on the uni t di sk? For exampl e consi der L
1
and L
?
(2) We have r epeatedl y used the par al l el ogr am l aw. Thi s hol ds i n a Hi l ber t
space. Does the par al l el ogr am l aw hol d i n spaces other than Hi l ber t
spaces?
(3) Convex sets, not necessari l y cl osed, do not have to have poi nts of
mi ni mum nor m.
(4) Cl osed sets, not necessari l y convex, do not have to have poi nts of
mi ni mum nor m. For exampl e,
C =
p=1

{ e
p
x where 1+ 1/p x < }
i s cl osed, not convex, and ther e i s not a poi nt of mi ni mum nor m.
Cor ollar y 12 I f M i s a cl osed l i near subspace of the Hi l ber t space {E, < , >},
x
0
i s i n E, and = i nf{| | x
0
- m| | : m M} then ther e i s one and onl y one y
0
i n M such that | | x
0
- y
0
| | = . Mor eover , <x
0
- y
0
, m> = 0 for al l m i n M.
I f y
1
M and y
1
y
0
then <x
0
- y
1
, m> 0 for al l m i n M.
Suggesti on of Proof: Let x
0
be i n E and y
0
be the cl osest poi nt i n M to x
0
.
Ther e i s such a poi nt because M i s cl osed and convex. We have fr om the
above theorem that
Re<x-y
0
, m-y
0
> 0 for al l m i n M. (*)
20
We hope to show that
Re<x-y
0
, m-y
0
> = 0 for al l m i n M.
Let u = 2y
0
- m. Note that u i s i n M. Hence,
0 Re<x-y
0
, u-y
0
> = Re< x-y
0
,y
0
-m>.
Thi s l ast i nequal i ty, together wi th i nequal i ty (*) above gi ves that
Re< x-y
0
, y
0
-m> = 0 for al l m i n M.
We now want to show that thi s equal i ty hol ds for the i magi nar y par t, too.
Choose v = y
0
+ i (m-y
0
). We have that
0 = Re<x-y
0
,i (m-y
0
)> = Re[ -i < x-y
0
, m-y
0
>
= I m <x-y
0
, m-y
0
>.
Ther efor e, < x-y
0
, m-y
0
> = 0 for m i n the l i near space.
Definit ion I f C i s a cl osed, convex set i n E then P
c
denotes the (possi bl y
nonl i near ) functi on fr om E to E such that i f x i s i n E then P
c
(x) i s the cl osest
el ement i n C to x.
Assign men t
(6.1) Compute <x,y> and | x+y|
2
- | x|
2
- | y|
2
for x = {0,i } and y = {0,1}.
(6.2) Suppose that {C
p
}
n
p=1
i s a col l ecti on of convex sets. Show that the
i nter secti on i s al so convex.
(6.3) Show that i f {x
p
}
n
p=1
i s a sequence of poi nts i n a convex set C and

1
n
t
p
= 1, wi th each t
p
0,
then
1
n
t
p
x
p
i s i n C.
(6.4) The cl osu re and the int erior of a convex set i s convex.
(6.5) I f E N M and M and N ar e cl osed subspaces wi th M N then there
i s z i n N, not i n M and wi th z per pendi cul ar to al l poi nts of M.
(Hi nt: Suppose that n i s i n N-M and m i s i n M. Let P(n) be the cl osest poi nt
to n that i s i n M and c = m + P(n). Then 0 = <n-P(n), c-P(n)>. Let z = n-P(n).)
(6.6) Consi der the space R
2
wi th the nor m gi ven by | | x| |
1
= | x
1
| + | x
2
| .
(a) Show that ever y cl osed set i n thi s space has a poi nt of mi ni mum
nor m.
(b) Show, by exampl e, that thi s may fai l to be uni que.
(c) What happens i f | | x| |

= max(| x
1
| , | x
2
| )?
21
(6.7) Note that x = (x
1
, x
2
, ... ) wher e x
p
=
(-1)
p
p
i s i n the r eal Hi l ber t space
(l i ttl e) RL
2
consi sti ng of squar e summabl e r eal sequences and C = {y: y
p
0,
y (l i ttl e) RL
2
} i s a convex set i n (l i ttl e) RL
2
. Fi nd the cl osest el ement i n C
to x.
(6.8) Gi ve an exampl e of a convex set C such that i f P
c
(x) i s the cl osest
el ement i n C to x, then P
c
(x) i s a functi on from E to E such that P
c
2
= P
c
but
P
c
i s not necessar i l y l i near . Can you char acter i ze those convex sets i n a
Hi l ber t space for whi ch the cl osest poi nt pr ojecti on i s a l i near functi on?
(6.9) Show that every cl osed, convex set i n a Hi l bert space has a poi nt of
mi ni mum nor m.
(6.10) Suppose that {
p
}
n
p=1
i s an or thonor mal sequence and x i s i n E. Show
that i f {a
p
}
n
p=1
i s a number sequence then
| x -
1
n
a
p
p
|
2
= | x|
2
+
1
n
| <x,
p
> - a
p
|
2
-
1
n
| <x,
p
>|
2
.
(Hi nt for seei ng the above equal i ty: expand the ri ght si de.)
How do you choose a
p
such that | x -
1
n
a
p
p
| i s mi ni mum? Let C = span
{
p
}
n
p=1
. Gi ve a for mul a for P
c
(x).
(Rema r k: Thi s pr obl em has sever al i nter esti ng par ts. Ther e i s the noti on of a cl osest
poi nt,
that
1
n
| <x,
p
>|
2
conver ges as n ,
that
1
n
<x,
p
>
p
conver ges i n E,
and that x =
1
n
<x,
p
>
p
i f {
p
} i s maxi mal or thonor mal .)
(6.11) Show that
n
(x) = si n(nx), n = 1, 2, ... i s an or thogonal sequence i n
L
2
[0,1]. Let C be the span of
n
(x), n = 1,2,..., 100. Gi ve a for mul a for P
c
(f)
where f i s gi ven by f(t) = t for 0 t 1.
(6.12) Show that
0
(x) = 1,
1
(x) = x,
2
(x) = (3x
2
- 1)/2,
3
(x) = (5x
3
- 3x)/2,
4
(x)
= (35x
4
- 30x
2
+ 3)/8 i s an or thogonal sequence i n L
2
[-1,1]. Let C be the span
of
n
(x), n = 0,1,...,4. Gi ve a for mul a for P
c
(f) wher e f(x) = x
3
+ x
2
+ x + 1.
22
(6.13) Let
L
n
(x) =
0
n
(-1)
p
n
p

1
p!
x
p
.
Compute L
0
thr ough L
3
. Let
n
(x) = e
-x
L
n
(x).
Show that {
n
(x)} i s an or thogonal sequence i n L
2
([0,), e
x
). Let C be the
span of
n
, n = 0,1,2,3. Gi ve a for mul a for P
c
(x) where f(x) = .25 i f 0 x 1,
= .75 i f 2 x 3, = 0 otherwi se.
MAP LE Rema r k: When one thi nks of the gr aph of cos(x/2) on the i nterval
[-1,1], one i s str uck by the r esembl ance of thi s gr aph to that of a quadr ati c,
tur ned down and tr ansl ated up, over the same i nter val . I t's cur i ous to
thi nk about how cl ose one mi ght appr oxi mate thi s tr anscandental functi on
wi th a quadr ati c functi on.
The fi r st thi ng to thi nk of i s the Tayl or Pol ynomi al of degr ee two.
> t ayl or ( cos( Pi * x/ 2) , x=0, 3) ;
> pl ot ( {cos( Pi * x/ 2) , 1 - Pi ^2* x^2/ 8}, x=- 1. . 1) ;
An al ter nate quadr ati c appr oxi mati on that i s mor e appr opr i ate to
thi s secti on i s to use the pol ynomi al s of pr obl em 6.12. These ar e cal l ed the
Legendr e pol ynomi al s. MAPLE knows these pol ynomi al s, and other s. Her e
ar e the fi r st thr ee Legendr e pol ynomi al s:
> wi t h( or t hopol y) :
> P( 0, x) ; P( 1, x) ; P( 2, x) ;
To get the best quadr ati c appr oxi mati on for cos(x/2) i n L
2
[0,1] we
compute the coeffi ci ents as i n pr obl em 6.10.
> a0: =i nt ( cos( Pi * x/ 2) * P( 0, x) , x=- 1. . 1) / i nt ( P( 0, x) ^2, x=- 1. . 1) ;
a1: =i nt ( cos( Pi * x/ 2) * P( 1, x) , x=- 1. . 1) / i nt ( P( 1, x) ^2, x=- 1. . 1) ;
a2: =i nt ( cos( Pi * x/ 2) * P( 2, x) , x=- 1. . 1) / i nt ( P( 2, x) ^2, x=- 1. . 1) ;
> pl ot ( {cos( Pi * x/ 2) , a0* P( 0, x) + a1* P( 1, x) + a2* P( 2, x) }, x=- 1. . 1) ;
To fur ther emphasi ze the i mpor tance of or thogonal pol ynomi al s, we
make one mor e i l l ustr ati on. Take n+1 poi nts chosen evenl y spaced on the
i nter val [-1,1] and choose the poi nt pai r s {x[i ],y[i ]} wher e y[i ] = f(x[i ]). Then
take the i nter pol ati ng pol ynomi al of degr ee n that fi ts the r esul ti ng n+1
poi nt-pai r s exactl y. Do you thi nk thi s pol ynomi al sequence mi ght conver ge
uni for ml y to f? Ther e i s a cl assi cal exampl e that i l l ustr ates how wr ong thi s
can be.
> f : =x- >1/ ( 1+25* x^2) ;
> f or i f r om 1 t o 9 do;
s[ i ] : = - 1 + ( i - 1) * 2/ 8;
f s[ i ] : = f ( s[ i ] ) ;
od;
23
> f i nt r p: =x- >i nt er p( [ s[ 1] , s[ 2] , s[ 3] , s[ 4] , s[ 5] , s[ 6] , s[ 7] , s[ 8] , s[ 9] ] ,
[ f s[ 1] , f s[ 2] , f s[ 3] , f s[ 4] , f s[ 5] , f s[ 6] , f s[ 7] , f s[ 8] , f s[ 9] ] , x) ;
> pl ot ( {f ( x) , f i nt r p( x) }, x=- 1. . 1) ;
The si tuati on -- l ack of cl oseness to f -- does not i mpr ove wi th mor e
poi nts; just modi fy thi s syntax to see.
24
Sect ion 7: Or t hogonal, Nonexpansive, & Self-Adjoint Pr oject ions
Some author s r equi r e that a pr ojecti on shoul d be nonexpansi ve,
or thogonal , sel f adjoi nt ... even l i near . These notes have made a poi nt that a
functi on wi th P
2
= P need not have these pr oper ti es. At l east geometr i cal l y,
the mi ni mum r equi r ement that P
2
= P defi nes a "pr ojecti on." So, one must
ask, what does one gai n wi th these other condi ti ons? Theor em 16 i n thi s
secti on pr ovi des an answer .
Lemma 13 (Pol ar i zati on I denti ty) I f T i s a l i near tr ansfor mati on then
<Tx, y> = <T(
x+y
2
),
x+y
2
> <T(
x-y
2
),
x-y
2
> +
i <T(
x+i y
2
),
x+i y
2
> i <T(
x-i y
2
),
x-i y
2
>.
Suggesti on of Pr oof: Just do i t!
Lemma 14 I f E i s a Hi l ber t space over the compl ex fi el d and T i s a l i near
functi on then these ar e equi val ent:
(a) <Tx, y> = <x, Ty> for each x and y i n E (or, T i s sel f adjoi nt), and
(b) <Tx, x> i s r eal for each x i n E.
Suggesti on of Pr oof: Use the Pol ar i zati on I denti ty.
Definit ion An Ort hogonal Project ion i s a pr ojecti on P for whi ch the nul l
space of P i s per pendi cul ar to the r ange of P.
Rema r k: Recal l that i f C of Theorem 11 i s a l i near subspace of E, then
statement (a) of Theorem 11 can be repl aced by
(a') z i s i n the subspace C and < x-z, c-z> = 0 for al l c i n C.
or
(a'') z i s i n the subspace C and <x-z, m> = 0 for al l m i n C.
Definit ion I f M i s a l i near space then M
= {n: <m,n> = 0 for al l m i n M},

R(P) i s the r ange of P, and and N(P) i s the nul l space of P.
Lemma 15 I f P i s a l i near , cl osest poi nt pr ojecti on onto a cl osed subspace M,
then (R(P))
= N(P) -- that i s, P i s an or thogonal pr ojecti on.

Suggesti on of Pr oof: Suppose that n N[P]. We want to show that <n,r> = 0
for al l r i n R(P). Si nce n N[P] then P(n) = 0. By Corol l ary 12, si nce P i s a
cl osest poi nt pr ojecti on, we have that for al l r i n R(P),
0 = <n-Pn,r-Pn> = <n-0, r-0> = <n,r>
for al l r i n R(P). Hence, nR(P)
.
Now, suppose that n R(P)
. We want to show that Pn=0.

25
Si nce n R(P)
, then <n,r> = 0 for al l r R, or <n-0,r-0> = 0 for al l r R.

But, thi s i s a char acter i zati on of cl osest poi nt.
EXAMP LE : Her e i s an exampl e of a pr ojecti on that i s not a cl osest poi nt:
P({x,y}) = {x-y,0}. What i s the r ange and nul l space of thi s pr ojecti on?
Theor em 16 Suppose that P i s a l i near pr ojecti on. These ar e equi val ent:
(a) P i s an or thogonal pr ojecti on.
(b) i f x i s i n E then <(1-P)x, Px> = 0.
(c) i f x and y are i n E, then <Px, y> = <x, Py>, i .e., P i s sel f adjoi nt.
(d) i f x and y ar e i n E then | x-Px |
2
+ | Px - Py |
2
= | x-Py |
2
, i .e., P i s a
r i ght tr i angl e pr ojecti on.
(e) i f x and y are i n E then | x-Px | | x-Py | , i .e., P i s a cl osest poi nt
pr ojecti on.
(f) <x-Px, z-Px > = 0 for al l x and for al l z i n P(E), i .e., P i s a r i ght angl e
pr ojecti on.
(g) | Px| | x| for each x i n E, i .e., P i s a non-expansi ve.
Suggesti on of Pr oofs:
ab Px i s i n the range of P and (1P) i s i n the nul l space of P.
bc Show <Px,x> i s r eal : <Px,x> = <Px,x-Px+Px> = | Px|
2
.
ca <n,r > = <n,Pr > = <Pn,r >.
ad Add to the l eft si de <x-Px,Px-Py> + <Px-Py,x-Px> = 0+0.
de Del ete | Px - Py|
2
from the l eft si de of d.
ef Use Corol l ary 12.
fb <x-Px, Px> = <x-Px, 0-Px> = 0 si nce 0 i s i n P(E)
cg | Px|
2
= <Px,Px> | Px| | x| .
gb Let y = Px + (x-Px). Note that Py = Px and | y-Py| = | | | x-Px| . By
(g) 0 | y|
2
- | Py|
2
= 2 Re * <Px,x-Px> + | |
2
| x-Px|
2
. I nvesti gate thi s
par abol a wi th real or equal to -i t, t real .
Rema r ks
(1) Some texts take al l projecti ons to be l i near. I t i s useful not to do thi s so
as to get or thogonal pr ojecti ons onto cl osed subsets.
(2) Some texts take a "resol uti on" of the i denti ty to be a sum of orthogonal
pr ojecti ons. We di d not do that because thi s di d not happen i n our
r esol uti on of matr i ces.
(3) One mi ght thi nk that i f A =
p=1

p
P
p
then these are equi val ent
(a) A i s sel f adjoi nt, and
(b) each
p
i s r eal .
26
Thi s di d not happen i n our exampl es. Not to wor r y! Ther e i s an i nner -
pr oduct wi th r espect to whi ch thi s i s a theor em.
Assign men t
(7.1) Gi ve exampl es whi ch contr ast pr ojecti ons, l i near pr ojecti ons and
or thogonal , l i near pr ojecti ons.
(7.2) Let C be the cl osed and bounded set {x: | x| 1 i n R
2
}. Let x = {1,1} and
P
c
(1,1) be the cl osest poi nt projecti on onto C. Fi nd y
1
i n C such that
<x-P
c
(x), y
1
- P
c
(x) > < 0.
Fi nd y
2
i n C such that
<x-P
c
(x), y
2
- P
c
(x) > = 0.
(7.3) Let S = L
2
[-1,1]. Let E = {f: fS, f(-x) = f(x)} and O = {f: fS, f(-x) = - f(x)}.
Show that E and O ar e or thogonal , l i near subspaces. Fi nd for mul as for
P
E
(f) and [1-P
E
](f). Show that P
E
i s a pr ojecti on, i s l i near , and i s
or thogonal .
MAP LE r ema r k: I n thi s MAPLE exer ci se, we constr uct a non-or thogonal
pr ojecti on of R
3
. To do thi s, choose u, v, and w l i near l y i ndependent. We
pr oject onto the subspace spanned by mul ti pl es of u. I nstead of bei ng an
or thogonal pr ojecti on, however , i t i s a pr ojecti on i n the "di r ecti on of v and
w." Thi s wi l l mean that
P(u) = u, P(v) = 0, and P(w) = 0.
Such a pr ojecti on can be accompl i shed as fol l ows:
Choose x i n R
3
. Wri te
x = au + bv + cw.
Defi ne P(x) = a u.
I n a si mi l ar manner , we get the pr ojecti on onto the subspace
spanned by v and i n the di r ecti on of u and w. And, we get the pr ojecti on
onto the supspace spanned by w i n the di r ecti on of u and v.
Her e's the techni que: choose, for exampl e, u = {1,0,1}, v = {1,1,0} and
w = {0, 0, 1}. We seek the projecti on of R
3
onto the space spanned by u al ong v
and w, the pr ojecti on of R
3
onto the space spanned by v al ong u and w, and
the pr ojecti on of R
3
onto spanned by w al ong u and v.
Take {x, y, z} to be a poi nt i n R
3
. we fi nd the fi r st matr i x, P
u
, whi ch i s
the projecti on onto u. Thus we seek a, b, and c so that a u + b v + c w =
{x,y,z} and we wi sh to wr i te a, b, and c as a functi on of x, y, and z.
Make the matr i x M such that the col umns ar e u, v, and w.
> wi t h( l i nal g) :
> u: =vect or ( [ 1, 0, 1] ) ; v: =vect or ( [ 1, 1, 0] ) ; w: =vect or ( [ 0, 0, 1] ) ;
> M: =t r anspose( ar r ay( [ [ u[ 1] , u[ 2] , u[ 3] ] ,
[ v[ 1] , v[ 2] , v[ 3] ] , [ w[ 1] , w[ 2] , w[ 3] ] ] ) ) ;
We know that {a, b, c} = M
-1
{x, y, z} and that the projecti on onto u
shoul d be
< M
-1
{x,y,z}, e
1
>.
27
> pr eP: =eval m( i nver se( M) &*vect or ( [ x, y, z] ) ) ;
> coef 1: =dot pr od( pr eP, vect or ( [ 1, 0, 0] ) ) ;
> pr oj 1: =eval m( coef 1*u) ;
> col 1: =vect or ( [ subs( {x=1, y=0, z=0}, pr oj 1[ 1] ) ,
subs( {x=1, y=0, z=0}, pr oj 1[ 2] ) ,
subs( {x=1, y=0, z=0}, pr oj 1[ 3] ) ] ) ;
subs( {x=0, y=1, z=0}, pr oj 1[ 2] ) ,
subs( {x=0, y=1, z=0}, pr oj 1[ 3] ) ] ) ;
subs( {x=0, y=0, z=1}, pr oj 1[ 2] ) ,
subs( {x=0, y=0, z=1}, pr oj 1[ 3] ) ] ) ;
> Pu: =eval m( t r anspose( mat r i x( [ col 1, col 2, col 3] ) ) ) ;
Her e i s a di ffer ent pr ojecti on.
> pr oj 2: =eval m( coef 2*v) ;
subs( {x=1, y=0, z=0}, pr oj 2[ 2] ) ,
subs( {x=1, y=0, z=0}, pr oj 2[ 3] ) ] ) ;
subs( {x=0, y=1, z=0}, pr oj 2[ 2] ) ,
subs( {x=0, y=1, z=0}, pr oj 2[ 3] ) ] ) ;
subs( {x=0, y=0, z=1}, pr oj 2[ 2] ) ,
subs( {x=0, y=0, z=1}, pr oj 2[ 3] ) ] ) ;
> Pv: =t r anspose( mat r i x( [ col 1, col 2, col 3] ) ) ;
And, her e i s the l ast pr ojecti on.
> pr oj 3: =eval m( coef 3*w) ;
subs( {x=1, y=0, z=0}, pr oj 3[ 2] ) ,
subs( {x=1, y=0, z=0}, pr oj 3[ 3] ) ] ) ;
subs( {x=0, y=1, z=0}, pr oj 3[ 2] ) ,
subs( {x=0, y=1, z=0}, pr oj 3[ 3] ) ] ) ;
subs( {x=0, y=0, z=1}, pr oj 3[ 2] ) ,
subs( {x=0, y=0, z=1}, pr oj 3[ 3] ) ] ) ;
> Pv: =t r anspose( mat r i x( [ col 1, col 2, col 3] ) ) ;
As, a check, i t shoul d be tr ue that the sum of the pr ojecti ons i s the i denti ty
matr i x.
28
Sect ion 8: Or t honor mal Vect or s
Surel y, any vi ew of a Hi l bert space wi l l have the noti on of
or thogonal i ty and or thogonal vector s at i ts cor e. The ver y noti on of
or thogonal i ty i s embedded i n the concept of the dot pr oduct. The cl assi cal
appl i cati ons of or thogonal functi ons per vade appl i ed mathemati cs. I ndeed,
ther e ar e books on or thogonal functi ons and thei r appl i cati on to appl i ed
mathemati cs.
We have seen that or thogonal vector s ar i se natur al l y as ei genvector s
of sel f adjoi nt tr ansfor mati ons. Thi s secti on pr ovi des a pr ocess for
gener ati ng an or thogonal set of vector s. I t al so exami nes the i mpl i cati on of
havi ng a maxi mal fami l y of or thogonal vector s.
Remar k One shoul d r evi ew the noti on of l i near l y i ndependent vector s and
ver i fy that any or thogonal col l ecti on i s l i near l y i ndependent.
Definit ion The col l ecti on {x
p
}
p=1
i s a maxi mal or thonor mal fami l y i f the
onl y vector y sati sfyi ng < x
p
, y > = 0, for al l p, i s y = 0.
Theor em 17 Let {x
p
}
p=1
be an orthonormal set i n {E, < , >}.

| <y, x
p
>|
2
| y|
2
,
wi th equal i ty hol di ng onl y i n case y =

<y, x
p
> x
p
.
Suggesti on of Pr oof: 0 | y -

<y, x
p
> x
p
|
2
= | y|
2
-

| <y, x
p
>|
2
.
Theor em 18 Suppose that {x
p
}
p=1
i s an or thonor mal sequence i n {E,< , >}.
These ar e equi val ent.
(a) {x
p
}
p=1
i s maxi mal - i n the sense that i f y i s i n E and <y,x
p
> = 0 for al l p
then y = 0.
(b) {x
p
}
p=1
i s an or thonor mal basi s - i n the sense that i f y i s i n E then
y =

<y,x
p
> x
p
.
(c) Par seval 's equal i ty hol ds - i f u and v ar e i n E then
29
< u,v> =
p

<u, x
p
> <x
p, V>.
(d) Bessel 's equal i ty hol ds - i f y i s i n E then
| y|
2
=

| <y, x
p
>|
2
.
(e) The span S of {x
p
}
p=1
i s dense i n E - i n the sense that i f y i s i n E then
ther e i s a sequence {u
p
}
p=1
i n S such that l i m
p
u
p
= y.
Suggesti on of Pr oof.
ab Suppose that {x
p
} i s maxi mal . Let M be al l y that can be wr i tten as

<y,x
p
> x
p
. Suppose M i s not E. Let z be i n M-E. Consi der
z
0
= z

<z,x
p
> x
p.
Then <z
0
,x
p
> = 0 for al l p.
bc <u,v> =
p
q
<u,x
p
> <x
q
,v> <x
p
,x
q
>.
cd <y,y> =
p
<y,x
p
> <x
p
,y>.
da I f <y,x
p
> = 0 for al l p then <y,y> =
p
<y,x
p
> <x
p
,y> = 0.
be I f u
n
=
p=1
n
<y,x
p
>x
p
then l i m
n
u
n
= y.
ea Suppose that y 0 and <y,x
p
> = 0 for al l p. Let {a
p
} be any number
sequence.
| y -

a
p
x
p
|
2
= | y|
2
+

| a
p
|
2
> | y|
2
.
Hence ther e i s no such {u
p
} for thi s y.
Definit ion I f {x
p
}
p=1
i s a l i near l y i ndependent sequence of vector s then the
Gramm-S chmi dt process gener ates an or thonor mal sequence as fol l ows:
u
1
= x
1,
v
1
= x
1
/| x
1
|
u
n
= x
n
-
p=1
n-1
<x
n
, v
p
> v
p
v
n
= u
n
/| u
n
| .
Remar k Thi s pr ocess gener ates an or thogonal sequence: i f n > k,
30
 = <x
n
-
p=1
n-1
<x
n
, v
p
>v
p
, v
k
> = 0.
I f 1 J n then the span of {x
p
}
J
p=1
i s the same as the span of {v
p
}
J
p=1
.
Theor em 19: Suppose that {P
n
}
n=1
i s a sequence of or thonor mal
pol ynomi al s, each havi ng degr ee n and that that the gener ati ng dot pr oduct
has the property that
<x f(x), g(x)> = <f(x), x g(x)>.
I t fol l ows that for each n, ther e ar e number s
n
,
n
, and
n
such that
P
n+1
(x) = (
n
x +
n
) P
n
(x) +
n
P
n-1
(x).
Wa r n in g: I n thi s di scusi on, P
n
i s a pol ynomi al , not a pr ojecti on.
Suggesti on for pr oof. Her e's why ther e i s the r ecur si on for mul a for
or thonor mal pol ynomi al s: We suppose that P
n
i s a pol ynomi al of degr ee n,
that = 0 i f n m and = 1 i f n = m, and that the dot product has the
property that <x f(x), g(x)> = <f(x), x g(x)>.
Choose
n
such that the functi on P
n+1
(x) -
n
x P
n
(x) i s a pol ynomi al
of degr ee n. Then ther e i s a sequence {
p
} of number s such that
P
n+1
(x) -
n
x P
n
(x) =
p=0
n

p
P
p
(x).
Suppose that k n-2.
0 = = <
n
x P
n
(x), P
k
(x)> + <
p=0
n

p
P
p
(x), P
k
(x)>
=
n
 +
k
=
n
 +
k
=

n
i =0
k+1

i
 +
k
= 0 +
k .
Hence, P
n+1
=
n
x P
n
+
n
P
n
(x) +
n-1
P
n-1
(x)
Examples Fr om the CRC:
Legendr e Pol ynomi al s: (n+1) P
n+1
(x) = (2n+1) x P
n
(x) - n P
n-1
(x).
Her e, <f, g> =
-1
1
f(x) g(x) dx.
Tschebysheff Pol ynomi al s: T
n+1
(x) = 2 x T
n
(x) - T
n-1
(x).
31
Her e, <f, g> =
-1
1
1x
2
f(x) g(x) dx.
Laguer r e Pol ynomi al s: (n+1) L
n+1
(x) = [2n+1-x] L
n
(x) - n L
n-1
(x).
Her e, <f, g> =
0

e
-x
f(x) g(x) dx.
Assign men t
(8.1) Let {x
p
}
p=1
be an or thonor mal sequence and {
p
}
p=1
be a number
sequence. These ar e equi val ent:
(a)
p=1
|
p
|
2
< , and
(b)
p=1

p
x
p
conver ges i n {E, < , >}.
(8.2) Let {x
p
}
N
p=1
be an or thonor mal sequence and M be the span of {x
p
}
N
p=1
.
Gi ve a for mul a for the cl osest poi nt pr ojecti on P
M
onto M.
(8.3) I f y i s i n E and {x
p
}
p=1
i s an or thonor mal sequence then the ser i es

p=1
<y, x
p
> x
p
conver ges.
(8.4) Gr amm-Schmi dt {e
1
, e
2
, e
1
+e
2
+e
3
} i n R
3
.
(8.5) Gr amm-Schmi dt {1,x,x
2
} i n L
2
([-1,1]).
(8.6) Gr amm-Schmi dt {e
-x
, xe
-x
, x
2
e
-x
} i n L
2
([0,),e
x
).
MAP LE Rema r k. The pr ocedur e to per for m the Gr am-Schmi dt pr ocess on
a sequence of vector s i s a par t of the MAPLE l i near al gebr a package. That
pr ocedur e uses the standar d dot-pr oduct. When usi ng thi s pr ocedur e, i t
shoul d be noted that the vector s r etur ned ar e not nor mal i zed.
> u: =vect or ( [ 1, 0, 1] ) ; v: =vect or ( [ 1, 1, 0] ) ; w: =vect or ( [ 0, 0, 1] ) ;
> Gr amSchmi dt ( {u, v, w}) ;
32
I t i s not so har d to wr i te a Gr am-Schmi dt pr ocedur e for dot-pr oducts
other than the standar d one, and even i n a functi on space.
MAPLE al so contai ns the standar d or thogonal functi ons. For
exampl e, her e ar e the Legendr e pol ynomi al s.
> wi t h( or t hopol y) ;
> P( 0, x) ; P( 1, x) ; P( 2, x) ; P( 3, x) ;
> i nt ( P( 2, x) * P( 3, x) , x=- 1. . 1) ;
33
Sect ion 9: The Finit e Dimension Par adigm
Si nce thi s col l ecti on of notes pur por ts to be about l i near oper ator s on
an i nner pr oduct space, one woul d hope to devel op an under standi ng of how
to make exampl es. A model that i s useful fol l ows the for m of the par adi gm
i ntr oduced i n thi s secti on. Two useful appl i cati ons of thi s par adi gm ar e
i ncl uded. We r etur n to fi ni te di mensi ons.
Theor em 20 I f A i s a sel f adjoi nt, l i near tr ansfor mati on fr om R
n
to R
n
then
ther e i s an or thonor mal sequence {
p
}
n
p=1
i n R
n
and a number sequence
{
p
}
n
p=1
such that i f x i s i n R
n
, then
Ax =
p=1
n

p
<x,
p
>
p.
Suggesti on of

Pr oof: Recal l that cor r espondi ng to any A ther e i s a sequence
of pr ojecti ons such that P
i
P
j
= 0 i f i j and such that x = P
i
x for al l x i n E.
Let M
i
= P
i
(E). Note that i f m M
i
then P
i
m = m and M
i
M
j
= 0 i f i j.
Sel ect i n each M
i
a maxi mal sequence of or thonor mal vector s and name the
mutual l y or thonor mal vector s i n the uni on of thi s col l ecti on for al l i {
1
,
2
,...,
k
}. We argue that k n, for R
n
has at most n l i near l y i ndependent
vectors. Even more, i t must be that k = n for otherwi se, l et y be
per pendi cul ar to thei r span. But, y =
i
P
i
y, each P
i
y i s i n M
i
, and the
vector s i n M
i
cannot al l be orthogonal to the the vectors i n {
p
}--other wi se,
the sequence {
p
} was not chosen as di rected. So, i t must be that {
p
} for ms
an or thonor mal basi s for R
n
. I f we now assume that A i s sel f adjoi nt, then
each P
i
i s a cl osest poi nt pr ojecti on onto M
i
. Thi s means that P
i
wi l l have a
r epr esentati on i n ter ms of the Fourier coefficient s
u = <u,
p
>
p
and A =
p
<u,
p
>
p.
Cor ollar y 21 Mor eover , i f m i s a posi ti ve i nteger , then
A
m
(x) =
p=1
n

p
m
<x,
p
>
p
.
Rema r ks
(1) Wi th A sel f adjoi nt,as i n Theor em 20, we have that N(A)
= R(A).
(2) Wi th A as supposed, i f
p
0 for 1 p k and
p
= 0 for k 
p
34
and v i s i n the r ange of A i f
v =
1
k
<v,
p
>
p
.
Definit ion I f A i s as i n Theor em 20 and m i s an i nteger such that m n and
i
0 for 1 i m and
i
= 0 i f i > m, then the general i zed i nverse i s a l i near
functi on such that
A(z) =
p=1
m

1
p
<z,
p
>
p
for each z i n R
n
.
Theor em 22 I f A i s as i n Theor em 20, then A

has the fol l owi ng pr oper ti es:
(1) AAA = A (2) AAA= A
(3) AA i s the or thogonal pr ojecti on onto R(A).
(4) AA i s the or thogonal pr ojecti on onto N(A)
.
Rema r k:
(1) One uses the gener al i zed i nver se to get a "best sol uti on" x for the
equati on Ax=y wher e A and y ar e gi ven and wher e ther e i s no x such that
Ax=y. Suppose that y R
n
and y R(A). Ther e i s no x such that Ax = y.
I n or der to fi nd v i n R(A) that i s cl osest to y, choose AAy. The above
establ i shes AA as the cl osest poi nt pr ojecti on onto the r ange of A. Now,
among al l u's such that Au = v, the one wi th smal l est nor m i s Ay. To see
thi s, fi r st note that AAy = v. I t r emai ns to show that Ay i s the cl osest
poi nt to zero i n the convex set of al l poi nts s that map to v, that i s, we need to
see that
< 0-Ay, s - Ay> = 0
for al l s such that As = v. To see thi s, note that s - Ay i s i n N(A), and the
above statement (4) shows that Ay =AA(Ay) i s per pendi cul ar to
ever ythi ng i n the nul l space of A.
(2) I f z i s i n R
n
, then AA z i s the cl osest poi nt u to z i n the r ange of A. Al so,
A(z) i s the el ement wi th the smal l est nor m among al l x's so that Ax = u.
Assign men t
(9.1) Compute A, AA, and AA for A =
0 2 0
2 0 0
0 0 0
.
(9.2) Let A =
-3/2 1/2 0
1/2-3/2 0
0 0 0
. Veri fy that {1,1,1} i s not i n the range of A. Fi nd v
i n the r ange of A so that | | v -{1,1,1}| | i s as smal l as possi bl e. Fi nd u so
that Au = v and u has the smal l est nor m among al l other x such that Ax =
v.
35
(9.3) The fol l owi ng matr i x i s sel f-adjoi nt, but A A
+
i s not di agonal :
4/3 1/3 -2/3

1/3 1/3 1/3
-2/3 1/3 4/3
.
MAP LE Rema r k: The methods suggested i n thi s secti on can be used wi th
MAPLE to get the fi ni te di mensi onal par adi gm r epr esentati on for a matr i x.
> A: = ar r ay( [ [ - 3/ 2, 1/ 2, 0] , [ 1/ 2, - 3/ 2, 0] , [ 0, 0, 0] ] ) ;
> ei genvect s( A) ;
> v1: =vect or ( [ - 1, 1, 0] ) ; v2: =vect or ( [ 1, 1, 0] ) ; v3: =vect or ( [ 0, 0, 1] ) ;
> u: =vect or ( [ x, y, z] ) ;
> r ep: = ( - 2) * i nner pr od( u, v1) * v1/ nor m( v1, 2) ^2 +
( - 1) * i nner pr od( u, v2) * v2/ nor m( v2, 2) ^2 +
( 0) * i nner pr od( u, v3) * v3/ nor m( v3, 2) ^2;
To under stand the next MAPLE command, note that i f u and v ar e vector s
and and ar e number s, then
add(u,v,,)
computes u + v.
> add( v1, v2, - x+y, - ( x+y) / 2) ;
As a check, i t shoul d be noted that the resul ts of the l ast l i ne are the
same as i f r egul ar matr i x mul ti pl i cati on had been per for med. The val ue of
thi s r epr esentati on has been expl ai ned i n thi s secti on.
36
Sect ion 10: Bounded, linear maps fr om E t o I C
The noti on of conti nui ty i s basi c to any anal ysi s of functi ons. I n a
study of Hi l ber t spaces, one shoul d have i n thei r r eper toi r e exampl es of
l i near functi ons that ar e conti nuous and exampl es that ar e not conti nuous.
We l ook fi r st at the si mpl est exampl es of conti nuous l i near functi ons on an
i nner pr oduct space: functi ons fr om E to the compl ex pl ane.
We pr ovi de a char acter i zati on of bounded, l i near functi ons fr om a
Hi l ber t space to the compl ex pl ane.
Theor em 23 Suppose that L i s a l i near functi on fr om {E,< , >} to I E. These
ar e equi val ent:
(a) Ther e i s a number b such that i f x i s i n E then | L(x)| b | x| - - L i s
bounded .
(b) i f {x
p
}
p=1
i s a sequence i n E wi th l i mi t y then l i m
p
L(x
p
) = L(y) - - L i s
cont i nuous.
Suggesti on of Pr oof:
ab | L(x
p
) - L(y)| = | L(x
p
-y)| b | x
p
-y| .
ba Suppose L i s conti nuous and ther e i s no such b. For each p, ther e i s z
p
wi th | L(z
p
)| p | z
p
| . Let x
p
=
z
p
| z
p
| p
. Then x
p
0, but
| L(x
p
)| =
| L(z
p
)|
| z
p
| p
p, so that | L(x
p
)| . Thi s contr adi cts the
assumpti on that L i s conti nuous.
Rema r k: The smal l est b i s the n orm of A. (Recal l the ear l i er di scussi on on
the nor m of a matr i x i n Chapter 4.)
Theor em 24 Suppose that L 0 i s a bounded l i near functi on and M = N(L).
Then M i s a cl osed, l i near subspace. Mor eover , i f L: E I C then di m M
=
1.
Suggesti on of Pr oof: Fi r st note that M i s a cl osed l i near subspace. Suppose
that L maps E to C. To prove that di m(M
) = 1, suppose that i t i s not. Let x

and y be l i near l y i ndependent member s of M
.
and consi der
z =
x
L(x)
-
y
L(y)
.
Si nce M

i s a subspace, z i s i n M
and L(z) = 0. Thus z MM

and
hence must be 0. Hence
x =
L(x)
L(y)
y
and x and y ar e not l i near l y i ndependent.
37
Theor em 25 (The Riesz Represent at ion Theorem)
Suppose that A i s a l i near functi on fr om {E, < , >} to I C. These ar e
equi val ent:
(a) Ther e i s y i n E such that A(x) = <x, y> for al l x i n E, and
(b) A i s bounded.
ab | Ax| = | <x,y>| | x| | y| .
ba By 24, N(A)
has di mensi on 1. Let z be i n N(A)
wi th | z| = 1. Let
y = A(z)* z.
Then,
<x,y> = <x, A(z)* z> = A(z) <x,z>
= A(z) {<Px, z> + <(1-P)x, z>} where P i s the cl osest poi nt projecti on onto
N(A)
= A(z) <(1-P)x,z> = A(z) <z,z> si nce N(A)

i s one di mensi onal
= A(z) = A(z) = A((1-P)x) = A(x)
Assign men t
(10.1) Let L be a functi on on L
2
defi ned by
L(x) =
p=1
| <x,e
p
>|
2
.
I s L a l i near functi on fr om L
2
to I C?
(10.2) Let L be defi ned on R
3
by L(x) = x
1
+ 3x
2
- 7x
3.
(a) Fi nd v such that L(x) = <x,v> for al l x i n R
3
.
(b) Fi nd b such that | L(x)| b | x| .
(10.3) Let A:L
2
[-1,1] I C by
A(f) =
0
1/2
f(x) dx.
Show that A i s conti nuous and fi nd g i n L
2
[-1,1] such that A(f) = <f, g>.
(10.4) Let C[-1,1] be the conti nuous functi ons i n L
2
[-1,1] and A: C[-1,1] I C
by A(f) = f(1/2). Show that A i s not conti nuous, that i s, show ther e i s a
sequence f
p
i n C[-1,1] wi th l i m
p
f
p
= 0, but l i m
p
A(f
p
) A(0). Show that
ther e i s no g i n L
2
[-1,1] such that
A(f) =
-1
1
f(x) g(x) dx.
(Hi nt: Recal l the sequence of functi ons {f
n
} that converges strongl y, but not
poi ntwi se.)
38
(10.5) For each f i n L
2
[0,1], l et Y(t) be a functi on such that y + 3 y = f, wi th
y(0) = 0. Defi ne A:L
2
[0,1]R by
A(f) =
0
1
y(t) dt.
Show that A i s a bounded l i near functi on on L
2
[0,1] and fi nd g such that
A(f) = < f, g >.
MAP LE Rema r k: Thi s i s an exer ci se i n getti ng the Ri esz r epr esentati on of
a conti nuous mappi ng fr om L
2
[0,1] to the compl ex pl ane.
Pi ck x i n [0,1]. We defi ne a poi nt - -a functi on -- G(x) i n L
2
[0,1]. Si nce
poi nts i n L
2
[0,1] ar e functi ons on [0,1], we wr i te thi s G as a functi on of not
onl y x, but al so t: G(x,t). The l i near mappi ng as defi ned by the Ri esz
theor em wi l l be
< G(x,t), f(t) > =
0
1
G(x,t) f(t) dt.
I t r emai ns to say what i s the l i near mappi ng for whi ch we wi l l get thi s
Ri esz r epr esentati on. Her e i s the defi ni ti on of L:
For each f, we defi ne L(f) = y wher e
y'' + 3 y + 2y = f, wi th y(0) = y(1) = 0.
We use MAPLE to cr eate thi s y. Then we make G(x, ).
Fi rst we sol ve the di fferenti al equati on wi th y(0) = 0. Thi s wi l l l eave
one constant to be eval uated.
> dsolve( {dif f ( y( x) ,x,x) +3*dif f ( y( x) ,x) +2*y( x) =f ( x) ,y( 0) =0},y( x) ) ;
> _ C2:=int ( ( exp( -1+u) -exp( -2*( 1-u) ) ) *f ( u) , u=0..1) / ( exp( -1) -exp( -2) ) ;
> ("");
The above makes G(x,u). I nput f, eval uate the i ntegr al s, and one has
a number . The Ri esz theor em says that we can fi nd a r epr esentati on i n
ter m of the dot pr oduct, whi ch i s an i ntegr al i n thi s case. Wi th a study of
those i ntegr al s, we can extr act G:
> G:=proc(x,u)
if u < x then
( exp( - x+u) - exp( - 2*( x- u) ) )
- ( exp( - 1+u) - exp( - 2+2*u) ) *exp( - x) / ( exp( - 1) - exp( - 2) )
+( exp( - 1+u) - exp( - 2+2*u) ) *exp( - 2*x) / ( exp( - 1) - exp( - 2) )
else - ( exp( - 1+u) - exp( - 2+2*u) ) *exp( - x) / ( exp( - 1) - exp( - 2) )
+( exp( - 1+u) - exp( - 2+2*u) ) *exp( - 2*x) / ( exp( - 1) - exp( - 2) )
f i
end;
Ther e ar e some pr oper ti es of Gr een's functi ons that may be fami l i ar . I t
sati sfi es the di fferenti al equati on for x u. We veri fy thi s:
39
> z1:=x->( exp( -x+u) -exp( -2*( x-u) ) )
+( exp( - 1+u) - exp( - 2+2*u) ) *exp( - 2*x) / ( exp( - 1) - exp( - 2) ) ;
> dif f ( z1( x) ,x,x) +3*dif f ( z1( x) ,x) +2*z1( x) ;
> z2:=x- >- ( exp( - 1+u) - exp( - 2+2*u) ) *exp( - x) / ( exp( - 1) - exp( - 2) )
+( exp( - 1+u) - exp( - 2+2*u) ) *exp( - 2*x) / ( exp( - 1) - exp( - 2) ) ;
> dif f ( z2( x) ,x,x) +3*dif f ( z2( x) ,x) +2*z2( x) ;
Ther e i s a symmetr y for the gr aph of G:
> plot3d(G,0..1,0..1);
0
0.200
0.400
0.600
0.800
1.00
0
0.200
0.400
0.600
0.800
1.00
-0.200
-0.150
-0.100
-0.0500
0
The obser vati on fr om the gr aph that the val ues of G ar e negati ve gi ves
i nfor mati on about the sol uti on for f's that have onl y non-negati ve val ues.
We compute a sol uti on for the boundar y val ue pr obl em wi th f the constant
functi on 1.
> w:=x- >int ( - ( exp( - 1+u) - exp( - 2+2*u) ) *exp( - x) / ( exp( - 1) - exp( - 2) )
+( exp( - 1+u) - exp( - 2+2*u) ) *exp( - 2*x) / ( exp( - 1) - exp( - 2) ) ,u=x..1)
+ int ( ( exp( - x+u) - exp( - 2*( x- u) ) )
+( exp( - 1+u) - exp( - 2+2*u) ) *exp( - 2*x) / ( exp( - 1) - exp( - 2) ) ,
u=0..x);
> w( 0) ;w( 1) ;dif f ( w( x) ,x,x) +3*dif f ( w( x) ,x) +2*w( x) -1;
> simplify(");
> simplify(");
> plot(w(x),x=0..1);
40
0 0. 2 0. 4 0. 6 0. 8 1
- 0. 14
- 0. 12
- 0. 1
- 0. 08
- 0. 06
- 0. 04
- 0. 02
0
x
41
Sect ion 11: Applicat ions t o Differ ent ial and Int egr al Equat ions
Thi s secti on wi l l pl ace some of the i deas that have come befor e i nto
the context of i ntegr al equati ons and i nto the context of or di nar y di ffer enti al
equati ons wi th boundar y condi ti ons. We have al r eady seen that i f A i s a
bounded, l i near tr ansfor mati on on E then exp(tA)(c) pr ovi des the sol uti on
for the i ni ti al val ue pr obl em
Y = AY, Y(0) = c.
Example
Let K(x,y) =
x(1-y) i f 0 x y 1
y(1-x) i f 0 y x 1
and l et A(f)(x) =
0
1
K(x,y) f(y) dy.
(1) A i s a bounded l i near tr ansfor mati on fr om L
2
[0,1] to L
2
[0,1] that i s sel f-
adjoi nt.
(2) I f f i s i n L
2
[0,1] then these are equi val ent:
(a) g(x) =
0
1
K(x,y) f(y) dy, so that g = A(f),
and
(b) g = -f wi th g(0) = g(1) = 0.
Defi ne B to have domai n the functi ons g i n L
2
[0,1] wi th g(0) = g(1) = 0 and
wi th two deri vati ves and B(g) = g. Note that A(f) = g i f and onl y i f f = B(g).
(3) These are equi val ent:
(a) i s a number , i s a functi on, and = A(), and
(b) =
1
n
2
2
and (x) = 2 si n(nx), for some posi ti ve i nteger n.
(4) I f f i s i n L
2
[0,1], then A(f) =

1
n
2
2
<f,
n
>
n
wher e
n
i s as above.
(5) (a) B(g) =
p=1
p
2
2
<g,
p
>
p
.
(b) The sol uti on for Z(t) + BZ(t) = 0, Z(0) = C i s
42
exp(-n
2

2
t) <C,
n
>
n
.
(6) The system Y = AY + F, E
0
Y(0) + E
1
Y(1) = 0 can be rewri tten as an
i ni ti al val ue pr obl em pr ovi ded [E
0
+ E
1
exp(A)] has an i nver se. I n any case,
a Green' s f unct i on G can be constr ucted wher e by, for appr opr i ate
functi ons F, sol uti ons for the system can be computed as
Y(x) =
0
1
G(x,t) F(t) dt.
Assign men t Fi nd matr i ces A, E
0
and E
1
such that the boundar y val ue
pr obl em
y'' + 3y' + 2y = f,
y(0) + y(1) = 0,
y'(0) - y'(1) = 0
i s wr i tten as a system Y = AY + F, E
0
Y(0) + E
1
Y(1) = 0.
Exa mple: Her e ar e gr aphs of the sol uti on y for
y'' + 3y' + 2y = f,
wi th y(0) = y(0) = 0, and another sol uti on wi th y(0) = 0, y(1) = 0, and another
sol uti on wi th y(0) - y(1) = 0, y'(0) + y'(1) = 0.
I s i t easy to i denti fy whi ch gr aph goes wi th whi ch sol uti on?
Fi gure 11.1
43
MAP LE Rema r k: I t i s i nter esti ng to thi nk of how the gr aphs of the
sol uti ons to
y'' + 3 y' + 2 y = 1
on [0,1] change as the boundar y condi ti ons change. At each ti me the
r el ati onshi p speci fi ed i n the di ffer enti al equati on between y and i ts
deri vati ves must hol d. How y starts at t = 0 i s modi fi ed i n order to sati sfy the
boundar y condi ti ons. Her e ar e thr ee exampl es.
> deq: =( D@@2) ( y) ( t ) + 3* D( y) ( t ) + 2* y( t ) = 1;
> bdr y1: =y( 0) = 0, D( y) ( 0) =0;
bdr y2: =y( 0) = 0, y( 1) = 0;
bdr y3: =y( 0) = y( 1) , D( y) ( 0) = - D( y) ( 1) ;
> y1: =dsol ve( {deq, bdr y1}, y( t ) ) ;
y2: =dsol ve( {deq, bdr y2}, y( t ) ) ;
y3: =dsol ve( {deq, bdr y3}, y( t ) ) ;
I t seems that MAPLE does not know how to sol ve thi s l ast boundar y
val ue pr obl em wi th boundar y condi ti ons bdr y3. Not to wor r y. Humans can
teach MAPLE!
> sol ve( { a+b = a* exp( - 1) +b* exp( - 2) ,
- a- 2* b = - ( - a* exp( - 1) - 2* b* exp( - 2) ) }, {a, b}) ;
> assi gn( " ) ;
> y3: =t - >1/ 2+a* exp( - t ) +b* exp( - 2* t ) ;
Her e's my check that y3 i s a sol uti on:
> ( D@@2) ( y3) ( t ) + 3* D( y3) ( t ) + 2* y3( t ) ;
> eval f ( y3( 0) - y3( 1) ) ;
> eval f ( D( y3) ( 0) +D( y3) ( 1) ) ;
Fi nal l y, her e i s a pl ot of the thr ee sol uti ons.
> pl ot ( {r hs( y1) , r hs( y2) , y3}, 0. . 1) ;
44
Sect ion 12: The Simple Par adigm for Linear Mappings
fr om E t o E
I t's ti me to put together two i deas: mi ght one dar e to guess the for m
for the par adi gm i n a gener al i nner pr oduct space? And, what i s speci al
about the par adi gm i n case the l i near functi on i t r epr esents i s a bounded
l i near mappi ng? The answer to both questi ons i s as coul d be expected.
Example Suppose that {
p
}
p=1
i s an i nfi ni te maxi mal or thonor mal
sequence and that {
p
}
p=1
i s a sequence of number s. The l i near
tr ansfor mati on A gi ven by
Ax =
p=1

p
<x,
p
>
p
i s bounded i f {
p
}
p=1
i s bounded and unbounded, other wi se. To ver i fy thi s,
we note that | Ax| sup
p
(|
p
| ) | x| for al l x i n E and that | A(
p
)| =
|
p
| |
p
| . The domai n of A i s al l x for whi ch
p=1
|
p
<x,
p
>|
2
< .
One can show that i f {
p
}
p=1
i s bounded then A has domai n al l of E and that
the domai n of A i s dense i n E i n any case. I t shoul d be noted that the
domai n of A
2
i s a (per haps pr oper ) subset of the domai n of A. (See the
Remar k bel ow.)
Mor eover , these ar e equi val ent:
(a) A has an i nver se, and
(b) each
p
0.
Fi nal l y, i f Bx =
p=1

p
<x,
p
>
p
,
then A and B commute.
Remar k To see that the domai n of A
2
i s a subset of the domai n of A, l et x be
i n the domai n of A
2
. Defi ne a sequence {
p
} by
p
=
p
i f |
p
| 1 and = 0 i f |
p
| < 1.
Suppose x i s i n D(A
2
). Then
45
>
p
|
p
2
<x,
p
>|
2

p
|
p
2
<x,
p
>|
2

p
|
p
<x,
p
>|
2
.
Then, | | x| |
2
+ | | A
2
x| |
2
=
p
| <x,
p
>|
2
+
p
|
p
2
<x,
p
>|
2

p
| <x,
p
>|
2
+
p
|
p
<x,
p
>|
2

p
|
p
<x,
p
>|
2
= | A(x)|
2
.
Assign men t
(12.1) Gi ve an exampl e of an unbounded l i near functi on whose i nver se i s
bounded.
(12.2) Gi ve an exampl e of an unbounded l i near functi on whose i nver se i s
unbounded.
(12.3) Gi ve the si mpl e par adi gm r epr esentati on for A:D L
2
[0,1] L
2
[0,1]
where D = {f: f exi sts and f(0) = f(1) = 0} and A(f) = f.
MAP LE r ema r k. We use MAPLE to veri fy that the sequence f
n
(x) =
si n(nx) for ms a set of ei genfuncti ons for A as defi ned i n Assi gnment 12.3.
We see that f
n
(0) = 0 and f
n
(1) = 0.
> di f f ( si n( n* Pi * x) , x, x) ;
The functi on A can be r epr esented i n thi s si mpl e par adi gm as we
saw i n Assi gnment 12.3. Wi th MAPLE's hel p do an exper i ment concer ni ng
A.
Consi der the l i near mappi ng
A(f) =
n=1
2 n
2

2
< f(t), si n(nt) > si n(nx).
We exami ne A( t (t-1) ).
> APr ox: =x- >2* sum( - n^2* Pi ^2* i nt ( t * ( t - 1) * si n( n* Pi * t ) , t =0. . 1) * si n( n* Pi * x) ,
n=1. . 10) ;
> pl ot ( APr ox( x) , x=0. . 1) ;
Thi s l ast was an appr oxi mati on for taki ng two der i vati ves of t (t-1).
Surel y, i t i s i rresi stabl e to take the i nverse of A, or at l east to approxi mate
thi s i nver se.
> APver se: =x- >2* sum( - i nt ( t * ( t - 1) * si n( n* Pi * t ) , t =0. . 1) /
( n^2* Pi ^2) * si n( n* Pi * x) , n=1. . 10) ;
> pl ot ( {x^4/ 12- x^3/ 6+x/ 12+. 0001, APver se( x) }, x=0. . 1) ;
The questi on i s, why di d APverse turn out to be
g(x) x
4
/12 - x
3
/6 + x/12
i nstead of, say x
4
/12 - x
3
/6? Two deri vati ves of both are zero....
46
Sect ion 13: Adjoint Oper at or s
Not ever y bounded l i near mappi ng i s sel f adjoi nt. You know thi s
fr om your exper i ence wi th matr i ces. Never mi nd. They al l have adjoi nts!
Suppose that A i s a bounded l i near tr ansfor mati on on E. Fi x z i n E
and defi ne L by L(x) = <Ax, z> for al l x i n E. L i s a bounded l i near functi on
from E to CI . By the Ri esz Theor em, ther e i s y i n E such that L(x) = <x,y> for
al l x i n E. Consi der thi s pai r i ng of z and y: for each z ther e i s onl y one y. We
defi ne the adjoi nt of A by pai r i ng thi s y wi th z.
Definit ion I f A i s a bounded l i near functi on on E then the adjoint of A, A*,
i s the functi on on E such that <Ax, z > = <x, A*z> for al l x i n the domai n of
A.
Obser vat ions
(1) A* i s a functi on fr om E to E that i s l i near and bounded.
(2) N(A*) = R(A)
(3) I f Ax =
p=1

p
<x,
p
>
p
for al l x then A*(y) =
p=1

p
*
<y,
p
>
p .
Theor em 26 Suppose that K i s a conti nuous functi on on [0,1]x[0,1] to R, so
that K(x,y) = K(y,x)*. Suppose al so that A i s defi ned by
A(f)(x) =
0
1
K(x,y)f(y) dy.
Then A i s sel f adjoi nt. Mor eover , i f M = l ub{<Ax, x>: | x| = 1} and m =
gl b{<Ax ,x>: | x| = 1} and i s an ei genval ue for A then m M.
Assign men t
(13.1) Suppose that K i s a conti nuous functi on from [0,1]x[0,1] to CI . I f
A(f)(x) =
0
1
K(x,y) f(y) dy
then A i s a bounded, l i near functi on fr om al l of L
2
[0,1]. Mor eover , A* i s
gi ven by
A*(g)(u) =
0
1
H(u,v) g(v) dv,
47
wher e H(u,v) = K(v,u)
*
, so that A i s sel f-adjoi nt i f K(x,y) = K(y,x)*.
(13.2) Let K(x,y) =
1
2
(y-x) -
1
4
i f x < y
1
2
(x-y) -
1
4
i f y < x
.
Defi ne A(f)(x) =
0
1
K(x,y) f(y) dy
(a) Show that A i s a bounded, l i near functi on fr om L
2
[0,1] to L
2
[0,1].
(b) Show that A* = A.
(c) Show that these are equi val ent:
(a) g = A(f)
(b) g = f, g(0) + g(1) = 0, g(0) + g(1) = 0.
(d) Fi nd the ei genval ues and ei genfuncti ons for A.
ans:
n
= - 1/[(2n+1) ]
2
, f
n
(x) = cos((2n+1) x) and si n((2n+1) x)
(e) Show that A has the property that i f {
p
}
p=1
i s the sequence of
ei genval ues then l i m
p
p
= 0.
MAP LE r ema r k: Li kel y, by now, the r eader i s awar e that what anal ysts cal l
the adjoi nt of an oper ator i s di ffer ent fr om what one often sees i n texts on
l i near al gebr a that i s cal l ed the adjoi nt of a matr i x. We per si st. But, we take
note that the l i near al gebr a package i n MAPLE fol l ows the notati on of the
l i near al gebr a texts, just as we fol l ow the pr ecedence of Hi l ber t space texts.
Per haps thi s MAPLE exer ci se wi l l gi ve under standi ng to the two
"adj oi nts".
> A: =ar r ay( [ [ 1, 2] , [ 3, 4] ] ) ;
> Aj oi nt : =adj oi nt ( A) ; Apose: =t r anspose( A) ;
> eval m( A &* Aj oi nt ) / det ( A) ;
> dot pr od( eval m( A &* vect or ( [ a, b] ) ) , vect or ( [ x, y] ) ) -
dot pr od( vect or ( [ a, b] ) , eval m( Apose &* vect or ( [ x, y] ) ) ) ;
> si mpl i f y( " ) ;
Thus, the tr anspose i s the adjoi nt of these notes. Her e i s a compl ex
exampl e.
> A: =ar r ay( [ [ 1+I , 2+3* I ] , [ 3- 2* I , 4] ] ) ;
> Aj oi nt : =adj oi nt ( A) ;
Apose: =t r anspose( A) ;
Aconj pose: =t r anspose( map( eval c, map( conj ugat e, A) ) ) ;
> eval m( A &* Aj oi nt ) / det ( A) ;
> dot pr od( eval m( A &* vect or ( [ a, b] ) ) , vect or ( [ x, y] ) )
- dot pr od( vect or ( [ a, b] ) , eval m( Aconj pose &* vect or ( [ x, y] ) ) ) ;
> si mpl i f y( " ) ;
48
Sect ion 14: Compact Set s
Dur i ng the next por ti on of the notes, we wi l l begi n an i nvesti gati on of
l i near , compact oper ator s. I n or der to do thi s study, ther e ar e sever al i deas
about number sets that shoul d be r emember ed: what i s a bounded number
set, what i s a compact number set, what i s a sequent ially compact number
set, and what i s a totally bounded number set. These noti ons car r y over to
nor med spaces, too
Definit ions A set S i s bounded i f ther e i s a number b such that i f x i s i n S
then | x| b.
A set C i s compact i f ever y open cover i ng of C has a fi ni te subcover i ng.
A set C i s sequent ially compact provi ded that i f {s
p
} i s a sequence wi th
val ues i n C then ther e i s a subsequence of {s
p
} that conver ges and has l i mi t
i n C.
A set S i s totally bounded i f, for each posi ti ve number c, ther e i s a fi ni te set
of poi nts {x
p
}
n
p=1
such that S i s contai ned i n
p
D
c
(x
p
). ( Her e, and i n the
r emai nder of the notes, D
c
(s) r epr esents the open di sk wi th center s and
r adi us c.)
Examples
(1) Unbounded sets are not total l y bounded.
(2) I n L
2
, cl (D
1
(0)) - - the cl osed uni t di sk - - i s not total l y bounded. Here's
why: Let c = 2/2. Si nce | | e
i
- e
j
| | = 2, then any col l ecti on of di sks of
r adi us l ess than 2/2 whi ch cover s al l of cl (D
1
(0)) must be i nfi ni te.
(3) I n R
3
, cl (D
1
(0)) i s total l y bounded. Here's why: Suppose that K i s a
posi ti ve i nteger . Choose poi nts F such that i f {a,b,c} i s one of them then
each of a, b, and c has the for m m/k wher e m i s an i nteger and -k m k.
For exampl e, i f k = 3, then {a,b,c} mi ght be {2/3, 1/3, -2/3}. I f {x,y,z} i s any
poi nt i n cl (D
1
(0)), then ther e i s a poi nt {a,b,c} i n F such that
| | {x,y,z} - {a,b,c}| |
2
<
3
k
2
.
Thus, i f c > 0, choose k such that

3
k
2
< c or
3
c
< k.
We cover cl (D
1
(0)) by di sks wi th radi us c and centers at the poi nts of F.
(4) (0,1) and [0,) ar e not sequenti al l y compact.
49
Rema r k: I n fi ni te di mensi ons, i f S i s a set, then these ar e equi val ent:
(a) S i s compact,
(b) S i s sequenti al l y compact, and
(c) S i s cl osed and bounded.
Our use of these i deas i s i l l ustr ated i n the next Theor em 27. To show the
gener al i ty, we r efer ence Int roduct ory Real Analysis, by Kol mogor ov and
Fomi n (tr ansl ated by Ri char d Si l ver man) and publ i shed by Dover . I n a
secti on number ed 11.2, they pr ove a Theor em 2: A metr i c space i s compact
i f and onl y i f i t i s total l y bounded and compl ete.
Theor em 27 I f S i s a subset of the Hi l bert space {E, < , >}, then these are
equi val ent:
(a) S i s sequenti al l y compact, and
(b) S i s cl osed and total l y bounded.
ab Prove S i s cl osed. Suppose l i m u
p
= v, each u
p
i s i n S. Sequenti al
compactness i mpl i es that v i s i n S.
To prove that S i s total l y bounded: Suppose > 0. Pi ck u
1
;

i f S i s not
contai ned i n D
(u
1
) pi ck u
2
i n S but not i n D
(u
1
). I f ther e ar e an i nfi ni te
number of such di sks then conti nue thi s pr ocess. Thi s pr oduces an i nfi ni te
sequence {u
p
} such that | u
p
-u
q
| > . S bei ng sequenti al l y compact i mpl i es
that thi s sequence has a subsequence that conver ges. Thi s i s a
contr adi cti on and ther e must be onl y a fi ni te number of such di sks.
ba Suppose that S i s total l y bounded and that {u
p
} i s an i nfi ni te sequence
wi th val ues i n S. A conver gent subsequence wi l l be extr acted. Choose a
fi ni te number of di sks wi th r adi us 1 that cover S. An i nfi ni te subsequence
of u l i es i n one of these. Cal l i t u
1
(p). Cover thi s di sk contai ni ng u
1
wi th a
fi ni te number of di sks wi th r adi us 1/2. An i nfi ni te subsequence of u
1
l i es i n
one of them, cal l thi s u
2
(p). ETC. Then | u
n
(n) - u
m
(m)| < 1/n for al l m > n.
Thi s i s a subsequence of u that conver ges. Si nce the space i s compl ete, the
sequence has a l i mi t. I f the set i s cl osed, the l i mi t i s i n S.
Remar k
I nsi ght i nto the geometr i c str uctur e of a sequenti al l y compact set i n L
2
i s
gai ned by r eal i zi ng that, whi l e a sequenti al l y compact set i n L
2
may be
i nfi ni te di mensi onal , i t can contai n no open set.
Assign men t
(1) Li st four poi nts {a,b} such that i f 0 < x < 1, 0 < y < 1, then
| | {x,y} - {a,b}| | <
1
2
for at l east one of the four {a,b}.
50
(2) Show that the Hi l bert cube i s total l y bounded. The Hi l bert cube consi st of
poi nts {x
p
} i n L
2
such that | x
n
|
1
2
n
.
(Hi nt: i f c > 0,
1
2
n
< c/2, and x i s i n the cube, then ther e i s x
n
and r
n
such
that x = x
n
+ r
n
, | r
n
| < c/2, and x
n
i s i n D
1
(0) R
n
. Thi s l ast set i s total l y
bounded.)
MAP LE r ema r k: The compact sets i n R
n
ar e pr eci sel y the cl osed and
bounded sets. That i s, a set i n R
n
i s compact i f and onl y i f i t i s cl osed and
bounded. I n a Hi l ber t space, ever y sequenti al l y compact set i s cl osed and
bounded, but there are cl osed and bounded sets that are not sequenti al l y
compact. I n fact, the cl osed uni t di sk i s such an exampl e. The fol l owi ng
syntax veri fi es that the sequence
2 si n(nx), n= 1, 2, 3, ...
i n an i nfi ni te sequence i n L
2
[0,1] i n the uni t di sk, any two ar e or thogonal ,
and the di stance between any two i s the square-root of two.
> i nt ( 2* si n( n* Pi * x) ^2, x=0. . 1) ;
> i nt ( 2* si n( n* Pi * x) * si n( m* Pi * x) , x=0. . 1) ;
> i nt ( 2* ( si n( n* Pi * x) - si n( m* Pi * x) ) ^2, x=0. . 1) ;
> si mpl i f y( " ) ;
> sqr t ( si mpl i f y( ( 4* n^3* m- 4* n* m^3) / ( n* m* ( n^2- m^2) ) ) ) ;
Note the r el ati onshi p between thi s exampl e and Exampl e 2 of thi s
secti on.
51
Sect ion 15: Compact Oper at or s
I f the si mpl e par adi gm for a l i near oper ator i n a Hi l ber t space i s to be
useful , we shoul d be abl e to l ook for char acter i sti cs of the par adi gm that
deter mi ne char acter i sti cs of the oper ator s. We now ask what condi ti ons on
the si mpl e par adi gm wi l l i mpl y that the i mage of the uni t di sk i s contai ned
i n a sequenti al l y compact set.
Theor em 28 Suppose that {
p
}
p=1
i s a maxi mal or thonor mal sequence i n E,
{
p
}
p=1
i s a sequence of number s, and
Ax=

p
<x,
p
>
p
.
These ar e equi val ent:
(1) l i m
p
|
p
| = 0, and
(2) cl (A( D
1
(0) )) i s sequenti al l y compact.
Suggesti on of Pr oof.
(1)(2). We show the cl osur e of A(D
1
(0)) i s total l y bounded. Si nce |
p
| 0
ther e i s a number B such that |
p
| |
2

2
| x|
2
.
We now pi ck out the set of poi nts requi red by the defi ni ti on of total l y
bounded: For 1 > c > 0, choose N such that max
Np
|
p
| < c/2. Si nce the
cl osed di sk wi th r adi us B i s a sequenti al l y compact subset of R
N
choose a
fi ni te set of poi nts {c
p
} i n R
N
such that i f i s i n R
N
wi th | | B then | -
c
p
|
2
< c/2 for some p. Now pi ck y i n the cl osur e of A(D
1
(0)). Ther e i s x i n
D
1
(0) such that Ax = y. Ther e i s p such that
|

q=1
N
[<y,
q
> - c
pq
]
2
| < c/2.
Al so, | y-c
p
| = |

q=1
N
[<y,
q
> - c
pq
]
2
| +
N+1
| <y,
q
>|
2
c/2 +
N+1
|
q
|
2
| <x,
q
>|
2
c/2 + c| x|
2
/2 = c.
(2)(1). Supppose l i m|
p
| 0. Ther e i s an i nfi ni te sequence {
n
} wi th |
n
|
= 1 and |
n
n
| c. Let x
n
be thi s sequence
n .
Then
52
| A(
n
) - A(
m
)|
2
= |
n
|
2
+ |
m
|
2
2 c
2
. Thus, A(x
n
) = A(
n
) has no
conver gent subsequence.
DEFINITION The l i near oper ator A i s compact i f A(D
1
(0)) i s contai ned i n a
cl osed and total l y bounded set.
Rema r ks.
(1) Note that the statement that the l i near operator i s compact i s equi val ent
to the fol l owi ng:
The l i near oper ator i s compact pr ovi ded i f {x
p
} i s a bounded sequence i n the
domai n of A, then ther e i s a subsequence {x[u
p
]} such that {A(x[u
p
])}
conver ges.
(2) I f {x
p
} i s bounded and conver ges weakl y to y and A i s a compact l i near
oper ator , then {A(x
p
)} conver ges str ongl y to A(y).
(3) Ever y compact l i near oper ator i s conti nuous.
Assign men t
(15.1) Let {
p
}
p=1
be a maxi mal or thonor mal sequence and defi ne A fr om E
to E by
Ax =

(2 -
1
p
) < x,
p
>
p
.
(a) Show that A i s sel f-adjoi nt, bounded, l i near , and not compact.
(b) Show that there i s no z such that Az = 2z.
(c) Show that | | Ax| | 2| | x| | for x i n E and i f b < 2 then there i s y i n E
such that | | Ay| | > b | | y| | .
(d) Sol ve the di fferenti al equati on Z = AZ, Z(0) = C E.
(15.2) Defi ne B from E to E by
Bx =

p <x,
p
>
p
.
(a) Show that B i s sel f adjoi nt, unbounded, l i near , and not compact.
(b) Show that there i s z such that Bz = 2z.
(c) Gi ve a poi nt x not i n the domai n of B.
(d) Let z be the sol uti on of the di fferenti al equati on Z + BZ = 0, Z(0) = C E.
Show that the sol uti on i s contracti ve i n the sense that | Z(t)| | C| .
(15.3) Sol ve these two equati ons
(a)
z
t
=
2
z
x
2
, z(t,0) = z(t,1) = 0, z(0,x) = f(x) L
2
[0,1]
53
(b)
z
t
=
2
z
x
2
+ g(t,x), z(t,0) = z(t,1) = 0, z(0,x) = f(x), wi th f and g(t, ) i n L
2
[0,1].
MAP LE r ema r k: We pr ovi de a sol uti on for the par ti al di ffer enti al equati on
z/t =
2
z/x
2
,
z(t,0) = z(t,1) = 0 for al l t > 0,
z(0,x) = x (x-1) for 0 < x < 1.
The methods we wi l l use empl oy the techni ques of thi s cour se. Thi s i s an
appr opr i ate pl ace to embed an under standi ng of the methods of sol vi ng
parti al di fferenti al equati ons by separati on of vari abl es i nto the context of
these notes.
Fi r st, we wr i te thi s par ti al di ffer enti al equati on as an or di nar y
di ffer enti al equati on i n L
2
[0,1]. Ther e Z' = AZ wher e Af = f'' wi th f(0) = f(1) =
0. We have seen the r epr esentati on for A i n Assi gnment 12.3:
A[f](x) =
n=1
2 n
2

2
< f(s), si n(ns) > si n(nx).
Mapl e computes the sol uti on and dr aws the sur face of the sol uti on.
> sum( 2* i nt ( s* ( s- 1) * si n( n* Pi * s) , s=0. . 1) * si n( n* Pi * x) , n=1. . 11) ;
> z: =( t , x) - >sum( exp( - n^2* Pi ^2* t ) * ( 2* i nt ( s* ( s- 1) * si n( n* Pi * s) , s=0. . 1) )
* si n( n* Pi * x) , n=1. . 11) ;
> pl ot 3d( z( t , x) , t =0. . 1/ 2, x=0. . 1, axes = NORMAL, or i ent at i on=[ 150, 65] ) ;
54
Sect ion 16: The Space of Bounded Linear Tr ansfor mat ions
One goal has been, and conti nues to be, the i nvesti gati on of
ei genval ue pr obl ems. That i s, we wi sh to establ i sh whether we can fi nd a
vector v and a number such that
Av = v.
We must do thi s i f we ar e to gener ate a par adi gm r epr esentati on for l i near
tr ansfor mati on A.
We wi l l see that i f A i s sel f-adjoi nt and compact, then ther e ar e
sol uti ons for an ei genval ue pr obl em. En r oute, we need to consi der the
operat or t opology. The noti on of a nor m for a matr i x has been suggested i n
these notes al r eady.
Definit ion BLT(E
1
,E
2
) i s the space of bounded, linear t ransformat ions from
E
1
to E
2
. I f L i s i n the space then | | L| |
1,2
i s the smal l est number b such
that | L(x)|
2
b | x|
1
for al l x i n E
1
.
Example
(1) Suppose that {
p
}
p=1
i s a maxi mal or thonor mal sequence i n E
1
and that
{
p
}
p=1
i s a maxi mal or thonor mal sequence i n E
2
. Defi ne
Lx =

(2 -
1
p
) <x,
p
>
p
.
I t fol l ows that | | L| |
1,2
= 2.
(2) Show that i f l i m
p
p
= 0 then l i m
n

p=1
n

p
< ,
p
>
p
conver ges i n BLT and
other wi se, l i m
n

p=1
n

p
<x,
p
>
p
strongl y -- i n the sense of Secti on 5.
(3) Let {
p
}
p=1
be a poi nt i n L
2
and {
p
}
p=1
be a maxi mal or thonor mal
sequence i n L
2
([0,1]). Defi ne K
n
(x,y) =
p=1
n

p
p
(x)
p
(y) on [0,1]x[0,1]. Then
0
1

0
1
| K
m
(x,y) - K
n
(x,y)|
2
dx dy =
n+1
m
|
p
|
2
.
55
Hence {K
p
}
p=1
i s Cauchy i n L
2
([0,1]x[0,1]) and has l i mi t
K(x,y) =
p=1
p
(x)
p
(y).
Defi ne K
n
fr om L
2
[0,1] to L
2
[0,1] by
K
n
(f)(x) =
0
1
K
n
(x,y) f(y) dy.
Then
0
1

0
1
[K
m
(x,y) - K
n
(x,y)] f(y) dy
2
dx

0
1

0
1
[K
m
(x,y) - K
n
(x,y)]
2
dy

0
1
| f(y)|
2
dy dx
| | K
m
K
n
| |
2
| | f| | .
Ther efor e, {K
p
}
p=1
i s Cauchy i n BLT on L
2
[0,1].
Theor em 29 Suppose that {L
p
}
p=1
i s a sequence of compact oper ator s i n
BLT and converges to K i n the nor m of BLT. Then K i s compact.
Suggesti on of Pr oof: Fi rst, note that K i s a bounded l i near oper ator because
cor r espondi ng to 1 > 0, ther e i s N such that i f n N then
| L
n
(x) - K(x)| 1 | x| ,
so that | K(x)| | L
n
(x)| + | K(x) - L
n
(x)| | L
n
(x)| + 1 | x| .
We now need to show that the cl osure of K(D
1
(0)) i s total l y bounded.
Let > 0. We seek a fi ni te number of di sks that wi l l cover the cl osur e of
K(D
1
(0)). Let N be such that i f n > N then | L
n
(x) - K(x)| < | x| /2. Si nce L
n
i s compact ther e i s a fi ni te sequence {c
p
} such that
p
D
/2
(c
p
) Cl (L
n
(D
1
(0))).
Let x D
1
(0). Then
| K(x) - c
p
| | K(x) - L
n
(x)| + | L
n
(x) - c
p
| .
Hence, Cl ( K(D
1
(0)))
p
D
/2
(c
p
).
Rema r k: We showed that i f l i m
p
|
p
| = 0 and {
p
} i s a maxi mal
or thonor mal sequence then
L
n
=
p=1
n

p
< ,
p
>
p
56
conver ges i n the BLT nor m to

p=1

p
< ,
p
>
p
.
Mor eover , each L
n
i s a compact oper ator . Thi s pr ovi des another exampl e of
a sequence of compact oper ator s that conver ges i n the BLT nor m and the
l i mi t i s compact.
Example Let K:[0,1]x[0,1] I C be conti nuous and l et
K(f)(x) =
0
1
K(x,y) f(y) dy.
Then K i s compact.
Theor em 30 I f A i s a bounded, l i near , sel f adjoi nt oper ator , then these ar e
equi val ent:
(1) b > 0 and | Ax| b | x| for al l x i n E, and
(2) b sup{ | <Ax,x>| : | x| = 1}
Proof: (1)(2) Thi s fol l ows fr om the Cauchy-Schwar tz i nequal i ty.
(2)(1) Let be the number | Ax| /| x| and z be the vector Ax/.
| Ax |
2
= <Ax,Ax> = < A(x), z >
= 1/4 [ <A(x+z), x+z> - <A(x-z), x-z> ]
1/4 b ( | x+z|
2
+ | x-z|
2
) and expandi ng these dot-pr oducts
= 1/2 b ( | x|
2
+ | z|
2
)
= 1/2 b[
2
| x|
2
+ 1/
2
| Ax|
2
] and usi ng the defi ni ti on of
= b | x| | Ax| .
Thus, | Ax| b| x| .
Exa mple: A par t of the hypothesi s for Theor em 30 i s that A i s sel f adjoi nt.
Thi s part of the hypothesi s was not used i n getti ng that (1) i mpl i es (2).
However, that part cannot be compl etel y removed, as i s i l l ustrated by the
fol l owi ng exampl e. Suppose that
1
and
2
ar e or thonor mal . Take
Ax = < x,
1
>
2
.
Then A i s l i near , and
| Ax| = | < x,
1
>| | x| .
However , | < Ax, x >|
2
= | < x ,
1
> < x ,
2
>|
2
and thi s i s str i ctl y l ess than
| < x ,
1
>|
2
+ | < x ,
2
>|
2
| x|
2
.
For exampl e, i f A(x,y) = (y,0), then A i s bounded, not sel f adjoi nt and
<A({x,y}), {x,y}> = x y.
57
The maxi mum val ue of x y subject to x
2
+ y
2
= 1 i s 1/2, not 1 -- whi ch i s
| | A| | .
Exa mples: Mi ndful that some author s take al l l i near pr ojecti ons to be
bounded (and or thogonal ) her e ar e two exampl es of l i near pr ojecti ons that
are unbounded: The second one has r ange a fi ni te di mensi oned space.
Exampl e (1) (Due to Mar y Chaml ee, Fal l ,'92) E = L
2
and
P({x
1
, x
2
, x
3
, ... , x
n
, ...}) = {x
1
- x
2
, 0, x
3
- 2 x
4
, 0, x
5
- 3 x
6
, 0, ... }.
Exampl e (2) (Due to Mahdi Zai dan, Fal l , '92) E = L
2
and
P({x
1
, x
2
, x
3
, ... , x
n
, ...}) = {x
1
+ 2x
3
+ 3x
4
+ ... , 0, 0, ... }.
Assign men t :
Suppose that K i s defi ned by
K(f)(x) =
0
1
cos((x-y)) f(y) dy.
Show that K has fi ni te di mensi onal r ange. Ar gue that i t i s a compact
oper ator . Compute | | K| | .
MAP LE r ema r ks. We have al r eady suggested that MAPLE can fi nd matr i x
nor ms. I n the next secti on, the concer n wi l l be i n getti ng ei genval ues for
sel f-adjoi nt l i near mappi ngs. A hi nt for how thi s wi l l go i s suggested i n
these MAPLE exper i ments.
> A: =ar r ay( [ [ 2, 1] , [ 1, 2] ] ) ;
> eval f ( nor m( A, 1) ) ; eval f ( nor m( A, 2) ) ; eval f ( nor m( A, i nf i ni t y) ) ;
> ei genval s( A) ;
> A: =ar r ay( [ [ 2, 1, 0] , [ 1, 0, 0] , [ 0, 0, 5] ] ) ;
> l ambda: =ei genval s( A) ;
57
Sect ion 17: The Eigenvalue Pr oblem
We now have the pr oper space i n whi ch to sol ve the ei genval ue
pr obl em: the space of the bounded l i near tr ansfor mati ons. We pr oceed to
show that the ei genval ue pr obl em can be sol ved.
Theor em 31 Suppose that A i s a compact, sel f-adjoi nt oper ator . Ther e i s a
r eal number and a non-zer o vector z such that | | = | | A| | and Az = z.
Pr oof: By Theor em 30, ther e i s a sequence {x
p
} i n E such that | x
p
| = 1 and
l i m
p
<Ax
p
,x
p
> = wher e | | = | | A| | . Si nce A i s compact, ther e i s z i n E
and a subsequence of x such that l i m
p
Ax
u(p)
= z. Then
l i m
p
<Ax
u(p)
, x
u(p)
> = .
0 | Ax
n
- x
n
|
2
= | Ax
n
|
2
+ | x
n
|
2
- <Ax
n
,x
n
> - <x
n
,Ax
n
>
(| | A| |
2
+ | |
2
) - 2 <Ax
n
,x
n
> = 2
2
- 2 <Ax
n
,x
n
> 0.
Thus, l i m
p
x
u(p)
= l i m
p
Ax
u(p)
= z,
or , 0 = A(z- l i m
p
x
u(p)
) = Az - A(l i m
p
x
u(p)
) = Az - z.
Theor em 32 I f A i s a compact, sel f adjoi nt oper ator and > 0, then the
number of ei genval ues wi th | | > i s fi ni te. Mor eover , i f i s an
ei genval ue for A and M = {x: Ax = x} then M i s fi ni te di mensi onal .
Pr oof: | Ax
p
- Ax
q
|
2
= |
p
x
p -
q
x
q
|
2
= |
p
|
2
| x
p
|
2
- |
q
|
2
| x
q
|
2
>
woul d l ead to a contr adi cti on i f ther e ar e i nfi ni tel y many and A i s compact.
Cor ollar y 33 I f A i s a compact, sel f-adjoi nt oper ator , then ther e i s a
sequence {
p
} of or thonor mal vector s and a number sequence {
p
} such that
A(x) =
p=1

p
<x,
p
>
p
.
Pr oof: We have establ i shed that A has an ei genval ue wi th | | = | | A| | .
Let M
1
= {x: Ax = x}. We al so know that M
1
i s fi ni te di mensi onal . Let {
p
}
N
1
be an or thonor mal basi s for M
1
. Then A(
p
) =
p
for each p. Let {
p
}
N+1
be
the addi ti onal or thonor mal vector s to make a maxi mal or thonor mal fami l y.
Let N be the (possi bl y i nfi ni te) combi nati ons of {
p
}
N+1
. Then A maps N i nto
N. To see thi s, i t suffi ces to see that A(
i
) i s i n N for each i . Suppose not.
Then
58
A(
i
) =
1
N
<A(
i
),
p
>
p
=
1
N

p
=
1
N

p
= 0.
But thi s i s i n N. Thus, A i s a compact, sel f adjoi nt oper ator on the Hi l ber t
space N. Fr om the above, i t has an ei genval ue wi th < | | A| | . We now
make M
2
and conti nue the pr ocess. The pr evi ous Theor em 32 establ i shes
that al l ei genval ues of A wi l l be found thi s away.
Remar k The above Theor em 31 r equi r es that A be compact and sel f-adjoi nt.
One cannot guar antee non-tr i vi al ei genval ues and ei genfuncti ons i f A i s
not sel f-adjoi nt. An exampl e fol l ows:
Let K(x,y) =
0 if 0 x y 1
e
y -x
- e
2(y-x)
i f 0 y x 1
and H(x,y) = K(y,x).
Her e i s the gr aph of K:
0
0.200
0.400
0.600
0.800
1.00
0
0.200
0.400
0.600
0.800
1.00
0
0.0500
0.100
0.150
0.200
Gr aph of K as gi ven above.
Defi ne A(f)(x) =
0
1
K(x,y) f(y) dy and B(g)(x) =
0
1
H(x,y) g(y) dy. Each of A
and B i s a bounded, compact l i near functi on fr om L
2
[0,1] to L
2
[0,1]. Al so A*
= B. Moreover, these are equi val ent:
(a) g = A(f), and
(b) g + 3g + 2g = f, g(0) = g(0) = 0.
Al so, these ar e equi val ent:
(c) g = B(f), and
(d) g - 3g + 2g = f, g(1) = g(1) = 0.
59
To see that A has no non-tr i vi al ei genval ues and ei genfuncti ons, note
that A(f)(x) =
0
x
K(x,y) f(y) dy, so that A i s a Vol ter r a I ntegr al Oper ator -- the
i ntegral i s from 0 to x, not 0 to 1. I f there were a number 0 and a
functi on f such that
f(x) =
0
x
K(x,y) f(y) dy
then | f(x)|
1
| |
n
B
n

x
n
n!
M, for each i nteger n, wher e | K(x,y)| B and
| f(x)| M. Hence | | f| | = 0.
Remar k Her e i s a pai r of sel f adjoi nt l i near oper ator s that ar e not compact:
A(x) = x for al l x i n any i nfi ni te di mensi onal Hi l ber t space, and B(f)(t) = t f(t)
i n L
2
([0,1]). A way to see the l atter one i s not compact i s to note that i t has
no ei genval ues.
Assign men t
(17.1) Go back and see wher e the pr oof of Theor em 31 uses A = A*.
(17.2) Suppose that E has a countabl e maxi mal or thonor mal sequence. I f P
i s a conti nuous pr ojecti on then these ar e equi val ent:
(a) P i s a compact oper ator , and
(b) The r ange of P i s fi ni te di mensi onal .
(17.3) Defi ni ti on: Suppose that A i s a bounded, sel f-adjoi nt operator on E and
that {
p
}
p=1
i s a maxi mal or thonor mal fami l y. Then tr aceA=

<A
n
,
n
>.
Thi s sum may be "". I f
n
i s a sequence of r eal number s, E i s L
2
[0,1] and
K(s,t) =

n

n
(s)
n
(t) and A(f)(x) =
0
1
K(x,y) f(y) dy
then tr aceA =
0
1
K(s,s) ds.
(17.4) Show that i f K(s,t) = 4 cos(s-t) then the operator gi ven by
A(f)(x) =
K(x,s) f(s) ds
i s a sel f-adjoi nt oper ator on L
2
[-,] wi th fi ni te di mensi onal r ange. Fi nd
pr ojecti ons and number s such that the oper ator i s gi ven by
60
1
P
1
+
2
P
2
+ +
n
P
n .
(17.5) Let {
p
}
p=1
be a maxi mal or thonor mal fami l y and A be defi ned by Ax
=
p=1

1
p
<x,
p
>
p+1
- so that A i s a "wei ghted shi ft". Show that
(a) | | A| |
2

p=1

1
p
2
.
(b) A = l i m
n
p=1
n

1
p
< ,
p
>
p+1
i n BLT.
(c) A i s compact.
(d) A has no nonzer o ei genval ues.
(e) Gi ve a for mul a for A*.
MAP LE r ema r ks
Her e i s the syntax for dr awi ng the gr aph of a ker nel . Seei ng the shape of
the gr aph gi ves a geometr i c under standi ng to the al gebr ai c noti on of
symmetr y: K(x,y) = K(y,x). The fol l owi ng gr aph ker nel i s not symmetr i c.
> K:=proc(x,y) if x <= y then 0
else exp( y-x) - exp( 2*( y-x) )
fi end;
> plot3d(K, 0..1, 0..1,axes=NORMAL);
Someti mes a fi ni te di mensi onal anal ogue of an i nfi ni te di mensi onal r esul t
wi l l show that the i deas ar e qui te easy i ndeed. Consi der thi s exampl e i n
connecti on wi th Assi gnment 17.5.
> A: =ar r ay( [ [ 0, 0, 0, 0] , [ 1, 0, 0, 0] , [ 0, 1/ 2, 0, 0] , [ 0, 0, 1/ 3, 0] ] ) ;
> ei genval s( A) ;
> t r anspose( A) ;
61
Sect ion 18: Nor mal Oper at or s and The Mor e Gener al Par adigm
Sel f adjoi nt oper ator s pl ay the r ol e i n BLT that r eal number s do i n I C -
- not onl y i n the sense that A = A*, but al so, i f T i s i n BLT, then there are
sel f-adjoi nt oper ator s A and B such that T = A + i B. I n fact A =
T+T*
2
and
B =
1
2i
(T-T*).
We conti nue our i nvesti gati on i nto the r epr esentati on of oper ator s as

p
P
p
. For r easons expl ai ned pr evi ousl y, we ask that the sequence of
pr ojecti ons shoul d for m a r esol uti on of the i denti ty.
Definit ion T i s normal provi ded that TT* = T*T.
Theor em 34 Suppose that T i s i n BLT and each of A and B i s sel f adjoi nt
wi th T = A + i B. These ar e equi val ent:
(a) T i s nor mal , and
(b) AB = BA.
Pr oof: To pr ove (b), si mpl y compute AB and BA. To pr ove (a), compute TT*
and T*T.
Theor em 35 Wi th the supposi ti on of Theor em 34, suppose al so that T i s
nor mal . These ar e equi val ent:
(a) T i s compact,
(b) Each of A and B i s compact, and
(c) T* i s compact.
Proof
(a)(b) | Tx|
2
= | (A+i B)x|
2
= | Ax|
2
+ i < Bx, Ax> - i <Ax, Bx > + | Bx|
2

| Ax |
2
| Bx|
2
(b)(a) | Tx
n
- Tx
m
|
2
= | Ax
n
- Ax
m
|
2
+ | Bx
n
- Bx
m
|
2
.
Theor em 36. Suppose that T i s a compact, nor mal oper ator . Then T has an
ei genval ue wi th max(| | A| | , | | B| | ) | | .
Pr oof: We know ther e i s an ei genval ue for A wi th | | = | | A| | . Consi der
{x: Ax = x} = N(I -A). Thi s i s a Hi l ber t space. To see that B maps N i nto
N, l et n N. Then ABn = BAn = Bn = Bn. Al so, B i s compact and
62
sel fadjoi nt on N. Ther e i s such that y = By. Let = +i and v = y. Then
Tv = (A+i B)v = Av +i Bv = v + i v = v. Al so, | | max(| | , | | ).
Theor em 37 I f T i s bounded and nor mal , then
(a) | Tx| = | T*x| ,
(b) I f Tx = x then T*x = *x, and
(c) I f , Tx = x, and Ty = y, then <x, y> = 0.
Remar k Actual l y, statement (a) i s equi val ent to the statement that T i s
nor mal .
Theor em 38 I f T i s a compact nor mal oper ator on E then ther e i s a fami l y
{
p
} of or thonor mal vector s whi ch i s maxi mal i n E and a sequence {
p
}
p=1
of compl ex number s such that, i f x i s i n E, then
Tx =
p=1

p
<x,
p
>
p
.
Mor eover , i f 0 i s not an ei genval ue of T, then the ei genvector s for m a
compl ete or thonor mal system.
The proof of thi s theorem i s just l i ke that of Theorem 36. The proof that the
or thonor mal fami l y i s maxi mal , or compl ete, can be ar gued as fol l ows:
suppose that v i s a nonzero vector wi th < v,
p
> = 0 for al l p. Then v must be
i n the nul l space of A so that Av= 0v. Thi s contr adi cts the supposi ti on that 0
i s not an ei genval ue of A.
Assign men t
(18.1). Suppose that {
p
}
p=1
i s a bounded sequence of compl ex number s and
{P
p
}
p=1
i s a sequence of or thogonal pr ojecti ons. Show that

p
P
p
i s
nor mal .
(18.2). Show that T(x) = <x,
p
>
p+1
i s not nor mal .
(18.3). Show that
1-1
0 0
i s a projecti on that i s not nor mal .
(18.4). Gi ve an exampl e of a matr i x that i s nor mal but not sel f adjoi nt.
MAP LE Remar k The chal l enge i s to make up a nor mal oper ator that i s not
sel f-adjoi nt.
> T: =ar r ay( [ [ 0, 0, 0, 0] , [ 1, 0, 0, 0] , [ 0, 1/ 2, 0, 0] , [ 0, 0, 1/ 3, 0] ] ) ;
> A: = eval m( A+t r anspose( A) ) / 2;
> B: = eval m( T- t r anspose( T) ) / ( 2* I ) ;
63
> eval m( A &* B) ; eval m( B &* A) ;
Si nce A and B do not commute, T i s not nor mal .
On the other hand her e i s a matr i x that i s nor mal , but not sel f adjoi nt:
T =
1 -1
1 1
.
> T:=array([[1,-1],[1,1]]);
> A:=evalm((T+transpose(T))/ 2);
> B:=evalm((T-transpose(T))/ (2*I));
Note that A and B ar e sel f-adjoi nt. Thi s i s cl ear for A. To see that i t i s tr ue
for B, compute:
> Bstar:=transpose(-B);
> evalm(A &* B); evalm(B &* A);
> evalm(A + I*B);
I s i t cl ear fr om thi s that T i s nor mal ?
64
Sect ion 19: Compact Oper at or s and Or t honor mal Families
A questi on ar i ses about how compact oper ator s map or thonor mal
fami l i es. I f T i s compact and nor mal and {
p
}
p=1
i s the sequence of
or thonor mal ei genval ues, then l i m
n
{T
n
}= l i m
n
{
n
}= 0. What i f T i s
compact, but not necessar i l y nor mal ( or sel f adjoi nt) and {
p
}
p=1
i s an
or thonor mal fami l y. Must l i m
n
{T
n
} = 0?
Theor em 39 Suppose that T i s compact and {
p
}
p=1
i s an or thonor mal
fami l y. Then l i m
n
{T
n
}= 0.
Pr oof: Suppose not. Ther e i s a subsequence {
u(p)
}
p=1
such that | T
u(p)
|
for some 0. Si nce {
u(p)
}
p=1
i s bounded and T i s compact, ther e i s a
subsequence of {T
u(p)
}
p=1
that conver ges and has l i mi t, say, v 0. Then,
l i m
n
< T
u(n),
v> = < v, v > 0.
But, al so,
l i m
n
< T
u(n),
v> = l i m
n
<
u(n)
, T*v > = 0
because these l ast ar e ter ms i n the Four i er expansi on of
T*v =
p
< T*v,
p
>
p
.
The ter ms of thi s sum must go to zer o. Thi s gets a contr adi cti on.
Remar k The i deas ar ound Theor em 38 pr ovi des an easy char acter i zati on of
when oper ator s commute.
Theor em 40 Supose that A and B ar e compact, nor mal oper ator s. These
ar e equi val ent:
(a) AB = BA, and
(b) Ther e i s a maxi mal or thonor mal fami l y {
p
}
p=1
whi ch ar e ei genvector s
for A and for B.
65
Pr oof. I f (b) hol ds, thi s i s cl ear fr om the r epr esentati on of Theor em 36.
Suppose that (a) hol ds and i s an ei genval ue of A. Let S be the subspace of
vector s x such that Ax = x. Because A and B commute, B maps S i nto S.
Hence ther e i s a sequence of ei genvector s for B that spans S. But, each of
these i s an ei genvector for A cor r espondi ng to . Thi s pr ocess i s symmetr i c
i n A and B. The r epr esentati on of Theor em 36 compl etes the r esul t.
Rema r ks
(1) i f i s a nonzer o ei genval ue for BA, then i t i s an ei genval ue for AB. To
see thi s, suppose BAx = x. Then
(AB) Ax = Ax = Ax.
Thus, i s a non zer o ei genval ue for AB.
2) Some use thi s l ast r esul t to char acter i ze nor mal oper ator s thi s way: an
oper ator i s nor mal i f and onl y i f i t and i ts adjoi nt can be "si mul taneousl y
di agonal i zed."
(3) Thi s cour se has i nvesti gated the r epr esentati on of l i near
tr ansfor mati ons as
p=1

p
<x,
p
>
p
. Thi s r epr esentati on gi ves i nsi ght as to
the natur e of l i near tr ansfor mati ons. We have found the r epr esentati on i s
appr opr i ate for compact, sel f-adjoi nt and nor mal oper ator s Fr om
exampl es, we have seen that i t gi ves an under standi ng to bounded, even i f
not compact oper ator s, and even to unbounded oper ator s on a Hi l ber t space.
The r epr esentati on shoul d gi ve i nsi ght and uni fi cati on to some of the i deas
that ar e encounter ed i n a study of i ntegr al equati on, Gr een's functi ons,
par ti al di ffer enti al equati ons, and Four i er ser i es.
Assign men t
(1) Fi nd the ei genval ues for
R(x) =
p=1
<x,
p
>
p+1
and for L(x) =
p=1
<x,
p+1
>
p.
(One of these has no ei genval ue and the other has ever y number i n the uni t
di sk as an ei genval ue.)
(2) Do the wei ghted l eft shi ft and ri ght shi ft have a representati on i n the
si mpl e par adi gm?
Ma ple Rema r k: The fi ni te di mensi onal anal ogue to the r i ght shi ft and the
l eft shi ft mi ght be expl or ed as fol l ows:
> R: =ar r ay( [ [ 0, 0, 0, 0] , [ 1, 0, 0, 0] , [ 0, 1, 0, 0] , [ 0, 0, 1, 0] ] ) ;
L: =ar r ay( [ [ 0, 1, 0, 0] , [ 0, 0, 1, 0] , [ 0, 0, 0, 1] , [ 0, 0, 0, 0] ] ) ;
66
Looked at thi s way, we see that ri ght-shi ft and l eft-shi ft are i nfi ni te
di mensi onal anal ogues of ni l potent oper ator s. We shoul d check to confi r m
what ar e the ei genval ues and ei genvector s of these two.
> ei genvect s( R) ; ei genvect s( L) ;
Ther e i s a new i dea that shoul d be br ought up her e. One coul d defi ne
"gener al i zed ei genvector s" for A as vector s v for whi ch ther e i s a number
such that
(A - I )v 0
but for whi ch
(A - I )
2
v = 0.
Thi s woul d be a gener al i zed ei genvector v of r ank 2. One coul d defi ne a
gener al i zed ei genvector of r ank k.
Her e's an exampl e:
> A: =ar r ay( [ [ 1, 1, 2] , [ 0, 1, 3] , [ 0, 0, 2] ] ) ;
> char pol y( A, x) ;
Note that 1 i s an ei genval ue of mul ti pl i ci ty 2, but has onl y one
ei genvector . We go l ooki ng for one gener al i zed ei genvector of r ank 2. To that
end, we fi nd the nul l space of (A-I )
2
:
> AmI dnt y: =eval m( ( A- di ag( 1, 1, 1) ) &* ( A- di ag( 1, 1, 1) ) ) ;
> nul l space( AmI dnt y) ;
Thus, we have that
(A-2 I ){5,3,1} = {0,0,0},
(A-1 I ){1,0,0} = {0,0,0},
and (A-1 I )
2
{0,1,0}={0,0,0}.
Sever al questi ons come to mi nd:
(1) What are the general i zed ei genvectors for R and L above?
(2) How do these fi t i nto the context and str uctur e for the par adi gm
pr esented i n these notes?
67
Sect ion 20: The Most Gener al Par adigm:
A Char act er izat ion of Compact Oper at or s
The par adi gm that has been sugested i n these notes i s appl i cabl e for
compact and nor mal oper ator s. Thi s i s a fai r l y sati sfactor y state of affai r s.
Yet, the si mpl e matr i ces
0 1
0 2
and
0 1 0
0 0 0
0 0 2

do not fi t i nto that si tuati on. We wi l l push the r epr esentati on one mor e
ti me. I n addi ti on to the sati sfacti on of havi ng a decomposi ton that i s
appl i cabl e to those two matri ces, we wi l l be abl e to obtai n the Fredhol m
Al ter nati ve Theor ems for mappi ngs wi th l ess hypothesi s.
Lemma 41 (1) I f A i s bounded and B i s compact then AB i s compact.
(2) I f A i s compact and B i s bounded then AB i s compact.
(3) I f T i s compact, then T
*
i s compact.
(Hi nt: si nce T i s compact, then TT
*
i s compact and <TT
*
x,x> = | T
*
x|
2
.)
Theor em 42. Suppose that T i s a compact operator from E to E. There are
maxi mal or thonor mal fami l i es {
p
} and {
p
} and a non-i ncr easi ng number
sequence {
p
}
p=1
such that l i m
p

p
= 0, and i f x i s i n E, then
T x =
p=1

p
< x ,
p
>
p.
Mor eover , the conver gence i s i n the nor m of BLT.
Pr oof:

Suppose that T i s compact. Fi r st, T* i s bounded si nce T i s. To see
thi s,
| T*x|
2
= <T*x,T*x> = <TT*x,x> | | T| | | T*x| | x| ,
so that | T*x| | | T| | | x| . Now, knowi ng that T* i s bounded and T i s
compact, we can get that T*T i s compact and i t i s sel fadjoi nt. Mor eover ,
<T*Tx, x> 0 so that al l the ei genval ues of T*T ar e nonnegati ve. Ar r ange
al l the ei genval ues i n decr easi ng or der . We have
T*Tx =
p

p
<x, x
p
> x
p
.
For each n such that
n
0, l et
y
n
= T(x
n
)/
n
.
Then < y
n
, y
m
> = <Tx
n
, Tx
m
>/
m

n
= <T*Tx
n
,x
m
>/
m

n
=
m

<x
n
,x
m
> = 0.
68
Thus, {y
p
} i s orthogonal , even orthonormal . Extend i t to be a m axi m al .
Then T(x
p
) =
p
y
p
even i f
p
= 0. Suppose that
x =
p
< x, x
p
> x
p
.
Tx = T(
p
< x, x
p
> x
p
) =
p
<x,x
p
> Tx
p
=
p

p
<x,x
p
> y
p
,
To see that thi s sum conver ges i n the BLT nor m,
|
n+1

p
<x,x
p
> y
p
|
2
=
p

p
| <x,x
p
>| <
p+1
| x|
2
Assign men t
(20.1) Per haps you wi l l agr ee that appl yi ng thi s decomposi ti on to the
matr i ces
0 1
0 2
and
0 1 0
0 0 0
0 0 2

i s i r r esi sti bl e. Note that thi s i s di ffer ent fr om the decomposi ti on whi ch we
had i n the fi r st of these notes.
(20.2) Wi th T as i n Theorem 42, What i s T*?
MAP LE Rema r k; We wi l l get the gener al i zed par adi gm for a matr i x T that
i s not nor mal .
> T: =ar r ay( [ [ 0, 1, 0] , [ 0, 0, 0] , [ 0, 0, 2] ] ) ;
We for m the sel f-adjoi nt matr i x A*A.
> A: = eval m( t r anspose( T) &* T) ;
> x[ 1] : =vect or ( [ 0, 1, 0] ) ; x[ 2] : = vect or ( [ 0, 0, 1] ) ; x[ 3] : = vect or ( [ 1, 0, 0] ) ;
> y[ 1] : =eval m( T &* x[ 1] / 1) ; y[ 2] : =eval m( T &* x[ 2] / 2) ; y[ 3] : =eval m( T &*
x[ 3] ) ;
> y[ 3] : =[ 0, 1, 0] ;
The proposal i s that T(u) = 1 <u,x
1
> y
1
+ 2 <u, x
2
> y
2
+ 0 <u,x
3
> y
3
.
We check thi s.
> u: =vect or ( [ a, b, c] ) ;
> eval m( T &* u) ;
> eval m( 1* i nner pr od( u, x[ 1] ) * y[ 1]
+ 2* i nner pr od( u, x[ 2] ) * y[ 2]
+ 0* i nner pr od( u, x[ 3] ) * y[ 3] ) ;
69
Sect ion 21: The Fr edholm Alt er nat ive Theor ems
The Fr edhol m Al ter nati ve theor ems concer n the equati on
(1-A)u = f.
These i deas come up r epeatedl y i n di ffer enti al equati ons and i n i ntegr al
equati ons. The Al ter nati ve Theor ems state necessar y and suffi ci ent
condi ti ons for the equati on (1-A)u = f to have a sol uti on u for some
pr evi ousl y speci fi ed f. Ther e ar e two al ter nati ves: ei ther the equati on has
exactl y one sol uti on for al l f or the equati on has many sol uti ons for some f's
and none for other s. Those for whi ch ther e ar e sol uti ons ar e char acter i zed.
Lemma 43. Suppose that A i s compact. Ther e ar e or thonor mal sequences
{
p
} and {
p
} such that [1-A] =
p=1
(1-
p
)< x,
p
>
p
.
Suggesti on of Pr oof: - A
*
- A + A
*
A i s compact and sel f adjoi nt. Al so,
[1-A]
*
[1-A] = 1 - A
*
- A + A
*
A.
Let - A
*
- A + A
*
A =
p

i
< ,
i
>
i
.
Al so, I =
p
< ,
i
>
i
so that [1-A]
*
[1-A] =
p
(1-
i
) < ,
i
>
i
.
Let
i
= 1- 1-
i
and
i
= [1-A](
i
)/(1-
i
).
Note that = 0.
Then x =
p
< x,
i
>
i
and [1-A](x) =
p=1
(1-
p
)< x,
p
>
p
.
Theor em 44 Suppose that A i s compact. Exactl y one of the fol l owi ng
al ter nati ves hol ds:
(a) i f f i s i n E then the equati on (1-A)u = f has onl y one sol uti on,
(b) the equati on (1-A)u = 0 has mor e than one sol uti on.
Theor em 45 I f A i s compact and the fi r st al ter nati ve hol ds for the equati on
(1-A)u = f then i t al so hol ds for the equati on (1-A*)u = f.
Theor em 46 Suppose A i s compact and the second al ter nati ve hol ds for the
equati on. Then these ar e equi val ent:
(1) the equati on (1-A)u = f has a sol uti on, and
(2) < f , z > = 0 for al l sol uti ons z for the equati on (1-A*)z = 0.
(Hi nt: to see that 2 i mpl i es 1, wri te out what i s the nul l space for (1-A*). )
70
Remar k For gener al l i near oper ator s, the questi on mi ght be, gi ven f, does
the equati on Bu = f have a sol uti on. For squar e matr i ces, we can compl etel y
answer the questi on:
Exactl y one of the fol l owi ng hol ds:
(a) det(B) 0 and i f f i s i n E then Bu = f has exactl y one sol uti on.
(b) det(B) = 0 and Bu = 0 has more than one sol uti on.
For the gener al compact oper ator , the al ter nati ves ar e not defi ni ti ve: Let
B =
p=1

1
p
< ,
p
>
p.
Then B(u) = 0 has onl y one sol uti on, but ther e ar e f's for whi ch B(u) = f has
no sol uti on.
Assign men t
(21.1) I n R
2
, l et A(x,y) = {x+y, 2y}. Show that (1-A)u = 0 and (1-A*)u = 0 have
non-tr i vi al sol uti ons. What condi ti ons must f sati sfy i n or der that (1-A)u = f
shoul d have a sol uti on?
(21.2) I n L
2
[0,1], l et A(f)(x) =
0
1
[1+cos(x) si n(t)] f(t) dt. Show that (1-A)u =
0 and (1-A*)u = 0 have non-tr i vi al sol uti ons. I f (1-A)u = f has a sol uti on,
what condi ti ons must f sati sfy?
(21.3) Show that i f A i s compact and nor mal then the r epr esentati on can be
[1-A] =
p=1
(1-
p
)< x,
p
>
p.
MAP LE Rema r k:
I n the matr i x si tuati on, the Fr edhol m Al ter nati ve Theor ems ar e no mor e
i nter esti ng than eval uati on of the deter mi nant. I f the deter mi nant i s not
zer o, the matr i x i s i n the fi r st al ter nati ve. I t the deter mi nant i s zer o, i t i s i n
the second al ter nati ve. Si nce
det(1-A) = det(1-A*),
i t i s no surpri se that the fi rst al ternati ve hol ds for 1-A i f and onl y i f i t hol ds
for 1-A*. The l ast of the Fr edhol m Al ter nati ve theor ems i s a l i ttl e mor e
i nter esti ng. We exami ne i t wi th a matr i x for whi ch det(1-A) = 0, so that the
second al ter nati ve hol ds.
> A: =ar r ay( [ [ 0, - 2, 1] , [ - 2, - 5, - 2] , [ - 3, - 4, 8] ] ) ;
> B: =eval m( di ag( 1, 1, 1) - A) ;
> det ( B) ;
71
Because det(1-A) = 0, thi s A i s i n the second al ter nati ve. To deter mi ne
for whi ch f's the equati on
(1-A)u = f
has a sol uti on u, we fi nd the nul l space of 1-A*.
> nul l space( eval m( di ag( 1, 1, 1) - t r anspose( A) ) ) ;
Ask what vector s ar e or thogonal to thi s one. Ther e i s a pl ane of them:
a [1,5,0] + b [1,0,5].
We sol ve (1-A)u = [1,5,0] and (1-A) = [1,0,5].
> sol ve( {x+2* y- z=1, 2* x+6* y+2* z=5, 3* x+4* y- 7* z=0}, {x, y, z}) ;
Fi nal l y, we pr edi ct the equati on (1-A)u = [1,5,1] has no sol uti on.
72
Sect ion 22: Closed Oper at or s
We change subjects now. The subject i s conti nui ty. The fi r st noti on
of conti nui ty encounter ed i n thi s cour se was appr oached fr om the
per specti ve of sequences and has been r estr i cted to l i near functi ons.
Bounded l i near functi ons and unbounded l i near functi ons ar e wor ds whi ch
wer e used to char acter i ze conti nuous and non-conti nuous l i near functi ons.
Later , a noti on was i nvesti gated that was mor e r estr i cti ve than conti unous.
I t i s that of a compact oper ator . Thus, we had these thr ee: compact
oper ator , bounded oper ator , and unbounded oper ator . Now we i ntr oduce
another that l i es among these. I t i s that of a cl osed oper ator .
Definit ion A functi on i s closed pr ovi ded i ts gr aph i s cl osed -- i n the sense
that i f {x
p
} has l i mi t y and {L(x
p
)} has l i mi t z then y i s i n the domai n of L
and L(y) = z.
Examples
(1) Let f(x) be the functi on from R to R defi ned by f(x) = 1/x i f x i s not zero
and f(0) = 0. Thi s functi on i s cl osed.
(2) Let f(x) be the functi on from R to R defi ned by f(x) = x/| x| i f x i s not zero
and f(0) = 0. Thi s functi on i s not cl osed.
We have char acter i zed maps whi ch sati sfy the par adi gm and whi ch
ar e compact, bounded, and unbounded dependi ng on the char acter of thei r
ei genval ues. Her e i s the si tuati on for cl osed oper ator s:
Theor em 47. Suppose that A =
p

p
< ,
p
>
p
and that the domai n i s as
l ar ge as possi bl e i n the sense that
D(A) = { x:
p
|
p
|
2
| < x,
p
>|
2
< }
I t fol l ows that the oper ator A i s cl osed.
Suggesti on of pr oof. Suppose that l i m
n
x
n
= y and l i m
n
A(x
n
) = z. We want to
show that
p
|
p
|
2
| < y,
p
>|
2
< and that
p

p
<y,
p
>
p
= z. For each N,
0 = < l i m
n
p

p
<x
n
,
p
>
p,
N
> - < z,
N
>
= l i m
n

N
< x
n
,
N
> - < z,
N
>
=
N
< y,
N
> - < z,
N
>.
Hence,
p
|
p
|
2
| < y,
p
>|
2
=
p
| < z,
p
>|
2
< ,
fr om whi ch we concl ude that y i s i n the domai n of A and z = A(y).
Remar k
(1) To see the gener al i ty of a cl osed oper ator , note that i f A i s a symmetr i c,
densel y defi ned oper ator i n a Hi l ber t space, then ther e i s a cl osed,
symmetr i c oper ator B such that A i s contai ned i n B.
(2) Because of the pr evi ous r emar k, we mi ght as wel l al ways take a
symmetr i c, densel y defi ned oper ator to be cl osed. Whi l e thi s wi l l not i mpl y
73
that the oper ator i s conti nuous, ther e i s thi s i dea: I t i s al ways possi bl e to
r edefi ne the i nner -pr oduct on D(A) such that the domai n of A becomes a
Hi l ber t space and A becomes a bounded oper ator on the domai n of A, I n
fact, for x and y i n the domai n of A, defi ne the new i nner -pr oduct by
< x, y>
n
= < x, y> + < Ax, Ay>.
Section 23: The Deficiency of A
We have examined the paradigm whereby A can be written as
(23.1) A(x) =
= p 1
p
< x ,
p
>
p
in case {
p
} -> 0 or in case {
p
} -> . We now examine the special case that {
1
p
} is bounded.
Remark: It goes without saying that we are supposing that no
p
is zero. It may be that {
p
} is not a maximal
orthonormal family. Then A could be extended to all of E by making the value zero at the extension of the
p
's
to a maximal orthonormal family. In that case, we can characterize the null space of this extension of A:
N(A) = { x: < x,
p
> = 0 if
p
0 }.
Theorem. Suppose that A can be written in the paradigm of (23.1) and that R(A) represents the range of
A. Then, these are equivalent:
(1) {
1
p
} is a bounded sequence,
(2) R(A) = {y: y =

p
0
< y ,
p
>
p
.
(3) R(A) = perp(N(A*)), where perp(N(A*)) is the collection of vectors perpendicular to the nullspace of A*.
(4) R(A) is closed.
Suggestions for a proof.
1 => 2. Suppose (1). Let
S = { y: y =

p
0
< y ,
p
>
p
}.
It is clear that the range of A is contained in S. To see that the range of A is S, suppose y is in S. Let
x =

p
0

1
p
< y,
p
>
p
.
Because of (1), x is in E. Also, Ax = y.
2 => 3. Recall that
74
A*(z) =
= p 1
p
* < z,
p
>
p
.
Also, as in the Remark above
(23.2) N(A*) = {z: < z,
p
> = 0 if
p
0 }.
And,
perp(N(A*)) = {w: < w,
p
> = 0 if
p
= 0}.
First we show that perp(N(A*)) is contained in R(A). Suppose that w is in perp(N(A*)). Then
w =
p
< w ,
p
>
p
=

p
0
< w,
p
>
p
.
To see that this w is in R(A), we need to produce x so that A(x) = w. As above, choose x =
x =

p
0

1
p
< y,
p
>
p
.
As before Ax = w. Thus, perp(N(A*)) is contained in R(A).
Finally, we show that R(A) is contained in perp(N(A*)). Suppose that y is in R(A) and z is in N(A*). Then y
is a combination of the
p
's
< y, z > = 
p
, z > =

p
0
< y,
p
> = 0 by equation (23.2).

Thus, y is in perp(N(A*)), completing the outline for 2 => 3.
(3) = (4) This follows because perp(N(A*)) is closed.
(4) => (1) Suppose that {1/
p
} is unbounded. Then
A
( ) 1
=
p
1
p
< z ,
p
>
p
is an unbounded operator. Consequently, A
( ) 1
cannot have domain all of E. Thus, there is z in E so that z is
not in the domain of A
( ) 1
or, what is the same, z is not in the R(A). Let y
n
be defined as follows:
y
n
=
= p 1
n
1
p
< z,
p
>
p
.
Then, A(y
n
) is in the range of A and is seen to be
A(y
n
) =
= p 1
n
1 < z,
p
>
p
.
75
Consequently, A(y
n
) is a sequence in R(A) that has limit z which is not in R(A). Thus, R(A) is not closed.
This is a contradiction.
This finishes an outline for a proof of the Theorem.
Definition. There are two numbers associated with A:
(A) = the dimension of the kernel of A.
and
(A) = the dimension of the kernel of A*.
We call the number (A) the deficiency of A. In a sense, this number measures how much the range of A is
deficient in filling E.
Examples:
1. If A is given by the simple paradigm,
A(x) =
= p 1
p
< x ,
p
>
p
,
then (A) is the number of
p
's such that
p
= 0, and (A) = (A).
2. If A is the right shift operator
A(x) =
= p 1
p
< x ,
p
>
+ p 1
,
so that
A*(x) =
= p 1
p
* < x ,
+ p 1
>
p
,
then (A) = 0 and (A) = 1.
3. We compute (A) and (A) for
K[f](x) = d
0
1
( ) cos ( ) x y ( ) f y y .
First, imagine K in the simple paradigm noting that the range of K is two dimensional and
K[cos( y)](x) =
( ) cos x
2
and K[sin( y)](x) =
( ) sin x
2
.
Hence, (A) in L
2
[0, 1] is . Also, (A) = , since K = K*.
76
4. We compute (K) for
K[f](x) = x d
0
x
( ) f y y - d
0
x
y ( ) f y y = d
0
x
( ) x y ( ) f y y
Suppose f is in the kernel of K. Then K[f](x) = 0. Taking the derivative with respect to x, we see that
0 =
x
K[f](x) = d
0
x
( ) f y y,
and
0 =
2
x
2
K[f](x) = f(x).
Thus, f is zero and the null space of A must be zero. Hence (A) = 0.
77
77
Sect ion 24: An Applicat ion: A Pr oblem in Cont r ol
We wi l l i l l ustr ate some of the i deas that we have encounter ed by
consi der i ng a contr ol pr obl em. To move gentl y to thi s pr obl em, we di scuss
four exampl es. The fi r st comes fr om 2
nd
year cal cul us. We wor k i t wi th
2
nd
year cal cul us techni ques and then note the di sadvantages of those
techni ques. The pr obl em i s r e-consi der ed and wor ked wi th mor e power ful
methods whi ch l ead to a method to sol ve the contr ol pr obl em.
Exa mple 1 Let M be the l i ne i n the i ntersecti on of the pl anes P
1
: x+y+z=0
and P
2
:x-y+z=0. Al so, consi der the poi nt {1,2,3} whi ch i s not i n ei ther pl ane.
The probl em i s to fi nd the cl osest poi nt i n M to {1,2,3}. Here i s a sol uti on
usi ng the techni ques of cal cul us:
P
1
i s the pl ane consi sti ng of poi nts {x, y, z} such that
<{x,y,z}, {1,1,1}> = 0.
P
2
i s the pl ane consi sti ng of poi nts {x, y, z} such that
<{x,y,z}, {1,-1,1}> = 0.
The l i ne M has the same di r ecti on as the vector
{1,1,1}{1,-1,1} = {2,0,-2} and contai ns {0,0,0}.
Hence, an equati on for thi s l i ne i s M(t) = {2t,0,-2t}. We want to choose t i n
order to mi ni mi ze | {2t,0,-2t} - {1,2,3}|
= (2t-1)
2
+ 2
2
+ (-2t-3)
2
= D(t).
Wel l , D(t) = 2(2t-1)2 + 2((-2t-3)(-2) = 16t + 8. And, D(t) = 0 provi ded that t =
-1/2. Hence, the cl osest poi nt i n M to {1,2,3} i s {-1,0,1}.
Remar k Thi s method onl y wor ks i n 3-D because the cr oss pr oduct i s uni que
to R
3
.
Example 2 Let L
1
and L
2
be l i near functi ons fr om E to I C. Let M be the
i nter secti on of the nul l space of L
1
and of L
2
. The pr obl em i s: gi ven u i n E,
fi nd the cl osest poi nt i n E to the nul l space of L
1
and L
2
, By the Ri esz
theor em, ther e i s y
1
and y
2
i n E such that L
1
= < , y
1
> and L
2
= < , y
2
>. I t
fol l ows that M = {x: <x, y
1
> = 0 = <x, y
2
> }. Then M = {x: <x, y
1
+ y
2
> = 0 for
al l and }. Al so, M
= span{y
1
, y
2
}. Now, suppose that u i s a poi nt of E
and P
m
i s the or thogonal pr ojecti on onto M and pr ovi des the cl osest poi nt i n
M to u. (Note u - P
m
(u) <u, y
1
>y
1
+ <u, y
2
>y
2
unl ess y
1
and y
2
are
or thonor mal .) We have u - P
m
(u) M
and P
m
(u) M. Fr om the fi r st of
these
u - P
m
(u) = y
1
+ y
2
for some and
and fr om the second
 = 0 = .
Hence,
<u, y
1
> = <y
1
, y
1
> + <y
2
, y
1
>
and <u, y
2
> = <y
1
, y
2
> + <y
2
, y
2
>,
or
78
( )
<u,y
1
> <u,y
2
>
=
<y
1
,y
1
> <y
2
,y
1
>
<y
1
,y
2
> <y
2
,y
2
>

.
We can sol ve thi s system uni quel y pr ovi ded
| y
1
|
2
| y
2
|
2
- | <y
1
,y
2
>|
2
0.
And, thi s i nequal i ty hol ds i f y
1
and y
2
ar e l i near l y i ndependent.
Assign men t
(1) Re-wor k exampl e 1 i n thi s context.
(2) Fi nd the poi nt i n the i nter secti on of the subspaces w+x+y+z = 0 and w-
x+y-z = 0 that i s cl osest to {0,1,2,3}.
Example 3 For many appl i cati ons, we make one fur ther var i ati on i n the
probl em. Let L
1
and L
2
be the l i near functi ons fr om E to I C. Let M = {x: L
1
(x)
= A
1
and L
2
(x) = A
2
}. Let L
1
= < , y
1
> and L
2
= < , y
2
>. Now, M i s a cl osed,
convex set. I f u E and i f we seek the cl osest poi nt i n M to u then P
m
(u)
agai n pr ovi des thi s poi nt. Whi l e P
m
(u) i s a pr ojecti on, i t i s not necessar i l y
l i near . We want a for mul a for P
m
(u). To get thi s for mul a, see that u -
P
m
(u) i s per pendi cul ar to N(L
1
)N(L
2
). We know that for al l y i n M, <u-
P
m
(u), y-P
m
(u)> 0. I n fact, i f n N(L
1
)N(L
2
) and y = n+P
m
(u) wi th y
M then <u-P
m
(u), n> 0. Al so, i f n N(L
1
)N(L
2
) then -n i s al so so that,
<u-P
m
(u), n> 0. Thus, u - P
m
(u) i s per pendi cul ar to the i nter secti on of
these two nul l spaces. As i n Exampl e 2, thi s i mpl i es that
u - P
m
(u) = y
1
+ y
2
.
We al so have the equati ons
A
1
= L
1
(P
m
(u)) = 
and A
2
= L
2
(P
m
(u)) = .
Suppose we know u and seek P
m
(u). From the above, we see that
u - y
1
- y
2
= P
m
(u). Hence,
<u, y
1
> - <y
1
, y
1
> - <y
2
, y
1
> = A
1,
<u, y
2
> - <y
1
, y
2
> - <y
2
, y
2
> = A
2.
Or ,
<u,y
1
> - A
1
<u,y
2
> - A
2
=
<y
1
,y
1
> <y
2
,y
1
>
<y
1
,y
2
> <y
2
,y
2
>

.
Example 4 Suppose that Q
0
R
2
and A i s a 2x2 matri x. Suppose that b R
2
and Q
1
R
2
. We seek v wi th mi ni mum nor m such that i f
Z = AZ + bv, Z(0) = Q
0
,
then al so, Z(1) = Q
1
. We know
Z(t) = exp(tA)Q
0
+
0
t
exp((t-s)A)bv(s) ds.
Si nce Q
1
= Z(1), We have
79
Q
1
= exp(A)Q
0
+
0
1
exp((1-s)A)bv(s) ds
or , Q
1
- exp(A)Q
0
=
0
1
exp((1-s)A)bv(s) ds.
Let
A
1
A
2
= Q
1
- exp(A)Q
0
. Let L
1
and L
2
be defi ned so that
L
1
(v)
L
2
(v)
=
0
1
exp((1-s)A)b v(s) ds .
To ask for the poi nt v wi th mi ni mum nor m whi ch sati sfi es A
1
=L
1
(v) and
A
2
= L
2
(v), we choose u i n the previ ous probl em to be 0. Let v be P
m
(0) wher e
P
m
i s the nonl i near pr ojecti on onto {x: L
1
(x) = A
1
and L
2
(x) = A
2
}. Fr om the
above exampl e,
v(s) = 0 - <exp((1-s)A)b, {1,0}> - <exp((1-s)A)b, {0,1}>
wher e and sati sfy an equati on such as that as that just pri or to thi s
exampl e.
Assign men t Fi nd v(s):[0,] R such that y + y = v, y(0) = 0, y(0) = 1, y() =
1, y() = 0.
Example Thi s exampl e shows that not ever y poi nt i s accessi bl e. Suppose
that P
1
and P
2
ar e or thogonal pr ojecti ons, A = P
1
+ P
2
, and that P
1
(b) = 0.
I f x i s i n E and u i s i n L
2
[0,1] then
| P
1
(x) -
0
1
exp((1-s)A)b u(s) ds| | P
1
(x)| .
Her e's why:
| P
1
(x) - (
0
1
exp((1-s))P
1
bu(s) ds +
0
1
exp((1-s))P
2
bu(s) ds)|
2
=
| P
1
(x) -
0
1
exp((1-s))u(s)ds P
2
b |
2
= | P
1
(x)|
2
+ |
0
1
exp((1-s))u(s) ds|
2
| P
2
(b)|
2
>_| P
1
(x)|
2
.
80
Definit ion I f A i s a l i near tr ansfor mati on, we denote by T the tr ajector y T =
{exp(sA)b: 0 s 1} and by .the accessi bl e poi nts = {z:
0
1
exp((1-s)A)b u(s)
ds, u L
2
[0,1]}.
Conject ur e:
= T
.
Section 25: Approximations in Reproducing Kernel Hilbert Spaces
In this section, we address two concepts. One is the wish that if {E, < , >} is an innerproduct space of real
valued functions on the interval [0,1], then there should be a function K from [0,1]x[0,1] to the real numbers,
such that if e is a number in [0,1], then g(x) = K(x,e) is in E, and if f is also in E, then
< f(x), K(x,e) > = f(e).
The second is the wish that if f is a reasonably nice function in E and
n
, n = 0 .. 5, is a sequence of functions
which have graphs that are triangles with peaks at t
p
= p/5, p=0..5, (See the next graphs.) then there should
be a reasonably good approximation for f(x) as
f(x) ~
= p 0
5
( ) f tp ( )
p
x .
In some circumstances these wishes come true.
Both these wishes are related to notions concerning use of the Dirac Delta function. The first idea is related
to the often used integral
d
0
1
( ) f x ( ) , x e x = f(e)
where ( ) , x e is the Dirac Delta. The second is the common notion that reasonably nice functions can be
written as sums of multiples of the Dirac Delta. We will show that the commonly used Dirac Delta cannot
exist in the Hilbert Space of square integrable functions on [0,1]. The formulation of such a function is
most often properly put in the context of distributions.
Definition: A Hilbert Space of functions on [0,1] has a reproducing kernel if there is a function K such that, for
each e in [0,1], the function g(x) = K(x,e) is in E and, for each f in E,
f(e) = < f(x), K(x,e) >.
81
Theorem. If { E, < , >} is a Hilbert Space of functions on [0,1] which has a reproducing kernel, then
normed convergence in E implies pointwise convergence on [0,1]. Moreover, if there is a number B such
that |K(x,x)| < B, then normed convergence in E implies uniform convergence on [0, 1].
Proof: A proof follows from this inequality: Suppose ( ) f
n
x , n=1, 2, 3,... is a sequence in E. Then
| ( ) f
n
e - ( ) f
m
e | = | < f
n
f
m
, K( ,e) > | | f
n
f
m
| |K( , e)| = | f
n
f
m
| ( ) K , e e .
Comments:
1. It follows that in the Hilbert Space of square integrable functions on [0, 1], if e is in [0,1], then g(x) = (x,e) is
not in E. This follows because we know that there are sequences of functions in that space which
converge in the norm of E, but do not converge pointwise on [0,1].
2. We will show later that the property of having normed convergence to imply pointwise convergence in
a Hilbert Space of functions on [0,1] is also a sufficient condition to assure that the Hilbert Space has a
reproducing kernel.
Example: Let E consist of all functions f on [0,1] that are continuous and for which the derivative f ' is
continuous except possibly for a finite number of jumps. Also, take f(0) to be zero for all functions in E.
Define the inner product by
< f, g > = d
0
1
f '(x) g '(x) x.
Take {E, < , >} to be the Hilbert Space formed from the completion of the innerproduct space defined
above. Define the function K by
K(x,y) = min(x,y).
Suppose f has a continuous derivative. Then
< f(x), K(x,y) > = d
0
y
x
f 1 x = f(y) - f(0) = f(y).
Consequently, this Hilbert Space has a reproducing kernel. It is the K as defined above.
Observation 1: With K as defined for this example, if a = 0.
To see this, note that if c < d, then the derivative of K(x,d) - K(x,c) is the characteristic function on [c, d].
That is, it is the function that is 1 on [c, d] and zero elsewhere. Thus, in the space of the example, the dot
product of
< K(x,b) - K(x,a), K(x,t) - K(x,s) >
is zero for non overlapping intervals.
82
Observation 2: If n is a positive integer, then the family { n (K(x, p/n) - K(x, (p-1)/n) }, p=1, 2, ... n, is an
orthonormal sequence in the Example above.
To see that this observation is true, the above Observation 1 establishes that the terms of the sequence are
orthogonal. We have only to see that they have norm 1. But, this is verified by integrating the constant
function 1 from (p-1)/n to p/n and multiply by n. The result is 1. For illustration, we draw the graph of one
of these functions in case n = 5.
Observation 3: Take
p
, p=1, 2, ...n, to be the orthonormal sequence defined above and suppose that f is in
the Hilbert Space of the Example. We know that the best approximation for f with the
p
's is given by
f ~
= p 1
n
c
p
p
and that the coefficients c
p
are the Fourier coefficients:
c
p
= < f ,
p
>.
Further, we know how good the approximation is
| f -
= p 1
n
c
p
p
|
2
f
2
-
= p 1
n
c
p
2
.
Observation 4: We calculate both terms of the right hand side of the above inequality in case f ' is
continuous except for a finite number f jumps:
f
2
= d
0
1
f '(x)
2
x,
and
83

= p 1
n
c
p
2
= 5
= p 1
n
d
p 1
n
p
n
f '(x) x
2
.
Observation 5: Recall that if f ' is continuous on the interval [(p-1)/n , p/n ] then
c
p
= f(p/n) - f((p-1)/n)
so that, in this case, we have an easy way to calculate the coefficients.
Two Approximation Examples:
Example 1. Suppose that f(x) = x
2
on [0,1]. We compute (a) f
2
, (b)
= p 1
n
c
p
2
, (c) the difference in these
two, and (d) draw a graph to illustrate the approximation.
Here is the computation for (a)
:= f x x
2
= d
0
1
d
d
x
( ) x
2
2
x
4
3
Here is the computation for (b)
= n d
p 1
n
p
n
d
d
x
( ) x
2
x
2 p 1
n
( ) / 3 2
=
= p 1
n
( ) 2 p 1
2
n
3
+ 1 4 n
2
3 n
2
We can make the computation of (c):
f
2
= 4/3,

= p 1
n
c
p
2
=
4
3
1
3 n
2
,
and | f -
= p 1
n
c
p
p
|
2

1
3 n
2
.
For (d) we draw the graphs with n = 5 to show the closeness of the approximation. The graph of f is black
and the approximation is red.
84
:= c
1
5
25
:= c
2
3 5
25
:= c
3
5
5
:= c
4
7 5
25
:= c
5
9 5
25
:= approx x
= p 1
n
c
p
5
K , x
p
n
K , x
p 1
n
Example 2. Suppose that f(x) = x ( ) 1 x on [0,1]. We make the same computations.
Here is the computation for (a)
= d
0
1
d
d
x
( ) x ( ) 1 x
2
x
1
3
Here is the computation for (b)
85
= n d
p 1
n
p
n
d
d
x
( ) x ( ) 1 x x
+ n 2 p 1
n
( ) / 3 2
=
= p 1
n
( ) + n 2 p 1
2
n
3
n
2
1
3 n
2
We can make the computation of (c):
f
2
= 1/3,

= p 1
n
c
p
2
=
1
3
1
3 n
2
,
and | f -
= p 1
n
c
p
p
|
2

1
3 n
2
.
For (d) we draw the graphs with n = 5 to show the closeness of the approximation. The graph of f is black
and the approximation is red.
:= c
1
4 5
25
:= c
2
2 5
25
:= c
3
0
:= c
4
2 5
25
:= c
5
4 5
25
:= approx x
= p 1
n
c
p
5
K , x
p
n
K , x
p 1
n
86
Alternate Computations.
We present here computations for the approximation that are simpler and make an interesting geometric
construction. Here is the idea. Suppose we have a function h which has a continuous derivative. Then the
Fourier Coefficients are easily computed:
d
a
b
h '(x) x = h(b) - h(a).
In this case, we can write the approximation as

= p 1
n
h
p
n
h
p 1
n
n
K , x
p
n
K , x
p 1
n
.
Observe that h(p/n) is multiplied by n (2 K(x,p/n) - K(x,(p-1)/n) - K(x,(p+1)/n)). Hence, an alternate
representation for the approximation that involves no computation of coefficients and the associate
integration can be created as follows:
define ( ) T
p
x = n (2 K(x,p/n) - K(x,(p-1)/n) - K(x,(p+1)/n)), p=1, 2, ...n.
Then h(x) ~
= p 1
n
h
p
n
( ) T
p
x
We present two illustrations.
Example 3: Let h(x) = sin(2 x). We draw the graphs using the approximation for h given above and with
n = 20.
:= approx x
= p 1
n
h
p
n
( ) T
p
x
87

Example 4: In this example, we make a graph of this modified representation for the approximation even
when h'(x) is not continuous. In this example, h is continuous, with a break in the derivative at x = 1/4. We
use n = 15 so that there is not a node point at the break.
:= approx x
= p 1
n
h
p
n
( ) T
p
x
In this example, it is interesting to see what are the supporting functions for the approximation.
88
Finally, we present the main theorem of this section.
Theorem. If { E, < , >} is a Hilbert Space of number valued functions on [0,1]. These are equivalent:
(1) There is a reproducing kernel for {E, < , >}.
(2) If x is in [0, 1] and L
x
is the function from E to the real number defined by
L
x
(f) = f(x)
for each f in E, then L
x
is continuous on E.
(3) Normed convergence in E implies pointwise convergence on [0, 1].
Indication of proof: It is routine to show that (2) and (3) are equivalent. We saw above that (1) implies (3).
To see that (2) implies (1), we use the Riesz Representation Theorem as follows. Suppose that x is in [0, 1]
and L
x
is defined as in (2). To assert that this is a continuous function from E to the numbers implies that
there is h in E such that
L
x
(f) = < f, h >.
Clearly, this Riesz point changes with x. Thus, we write, for each f in E,
f(x) = L
x
(f) = < f, h
x
> = < f(t), h(t,x) >.
89
81
I n d ex: Hi lber t Sp a ce Not es
Parseval 's i nequal i ty 27
Adjoi nt 45-47 pol ari zati on 23
Banach Space 15 pol ynomi al 29-30
Bessel 's i nequal i ty 28 projecti on, defi ni ti on 1
BLT(E
1
,E
2
) 53 projecti on, cl osest poi nt 18-22,
32-34, 77-78 Cayl ey-Hami l ton 2
projecti on, ri ght tri angl e 24 Cauchy-Schwartz 7
projecti on, unbounded, 56 cl osed, operator 72-73
orthogonal 23 cl osed, set 18-20, 35, 47-50
Rei sz Representati on Theorem 36,
37
cl osest poi nt 18-24
cl osure 20, 48
resol uti on of i denti ty 59 compact, operator 49, 50-71
sel f-adjoi nt 7, 23, 45 compact, set 47-48
separabl e compl ete space 15
Shi ft operators 65 conti nuous 35
spectrum control 77
total l y bounded 47 convergence of operators 53
convergence, poi ntwi se 14 Tschebysheff 29
unbounded operator 65, 72 convergence, strong 14
convergence, uni form 14
convergence, weak 14
convex set 18
di fferenti al equati ons 5, 8, 40-42,
58
ei genval ue -vector 8-12, 27, 32,
43-45, 53-67
Fouri er 15, 31, 65
Fredhol m al ternati ve 69-71
general i zed ei genvector 66
general i zed i nverse 33
Gerschgori n ci rcl e theorem 11-12
Gramm-Schmi dt 28-30
Green's functi ons 41, 65
Green, Wi l l i am F. 81
Hi l bert space 15
I nnerproduct space 7
i nteri or 20
Jordan Form 1
Laguerre 29
Legendre 30
maxi mal orthonormal fami l y 27
matri x norms 12
ni l potent 1, 5, 66
non-expansi ve 1, 23
normal operator 61-62, 67
operator topol ogy 53
Paral l el ogram 18
82
A Comp en d i u m
of P r oblems

Herod

Încărcat de

Informații document

Descriere originală:

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

Herod

Încărcat de

Drepturi de autor:

Formate disponibile

Ma t h 6580

conver ges for | u

= {n: <m,n> = 0 for al l m i n M},

= N(P) -- that i s, P i s an or thogonal pr ojecti on.

. We want to show that Pn=0.

, then <n,r> = 0 for al l r R, or <n-0,r-0> = 0 for al l r R.

4/3 1/3 -2/3

) = 1, suppose that i t i s not. Let x

and L(z) = 0. Thus z MM

has di mensi on 1. Let z be i n N(A)

S-ar putea să vă placă și